xref: /sqlite-3.40.0/src/where.c (revision 9bccde3d)
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.
10 **
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".
18 */
19 #include "sqliteInt.h"
20 #include "whereInt.h"
21 
22 /* Forward declaration of methods */
23 static int whereLoopResize(sqlite3*, WhereLoop*, int);
24 
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
29 
30 
31 /*
32 ** Return the estimated number of output rows from a WHERE clause
33 */
34 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){
35   return pWInfo->nRowOut;
36 }
37 
38 /*
39 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this
40 ** WHERE clause returns outputs for DISTINCT processing.
41 */
42 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){
43   return pWInfo->eDistinct;
44 }
45 
46 /*
47 ** Return TRUE if the WHERE clause returns rows in ORDER BY order.
48 ** Return FALSE if the output needs to be sorted.
49 */
50 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){
51   return pWInfo->nOBSat;
52 }
53 
54 /*
55 ** Return the VDBE address or label to jump to in order to continue
56 ** immediately with the next row of a WHERE clause.
57 */
58 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){
59   assert( pWInfo->iContinue!=0 );
60   return pWInfo->iContinue;
61 }
62 
63 /*
64 ** Return the VDBE address or label to jump to in order to break
65 ** out of a WHERE loop.
66 */
67 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){
68   return pWInfo->iBreak;
69 }
70 
71 /*
72 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to
73 ** operate directly on the rowis returned by a WHERE clause.  Return
74 ** ONEPASS_SINGLE (1) if the statement can operation directly because only
75 ** a single row is to be changed.  Return ONEPASS_MULTI (2) if the one-pass
76 ** optimization can be used on multiple
77 **
78 ** If the ONEPASS optimization is used (if this routine returns true)
79 ** then also write the indices of open cursors used by ONEPASS
80 ** into aiCur[0] and aiCur[1].  iaCur[0] gets the cursor of the data
81 ** table and iaCur[1] gets the cursor used by an auxiliary index.
82 ** Either value may be -1, indicating that cursor is not used.
83 ** Any cursors returned will have been opened for writing.
84 **
85 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is
86 ** unable to use the ONEPASS optimization.
87 */
88 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){
89   memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2);
90 #ifdef WHERETRACE_ENABLED
91   if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){
92     sqlite3DebugPrintf("%s cursors: %d %d\n",
93          pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI",
94          aiCur[0], aiCur[1]);
95   }
96 #endif
97   return pWInfo->eOnePass;
98 }
99 
100 /*
101 ** Move the content of pSrc into pDest
102 */
103 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){
104   pDest->n = pSrc->n;
105   memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0]));
106 }
107 
108 /*
109 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet.
110 **
111 ** The new entry might overwrite an existing entry, or it might be
112 ** appended, or it might be discarded.  Do whatever is the right thing
113 ** so that pSet keeps the N_OR_COST best entries seen so far.
114 */
115 static int whereOrInsert(
116   WhereOrSet *pSet,      /* The WhereOrSet to be updated */
117   Bitmask prereq,        /* Prerequisites of the new entry */
118   LogEst rRun,           /* Run-cost of the new entry */
119   LogEst nOut            /* Number of outputs for the new entry */
120 ){
121   u16 i;
122   WhereOrCost *p;
123   for(i=pSet->n, p=pSet->a; i>0; i--, p++){
124     if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){
125       goto whereOrInsert_done;
126     }
127     if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){
128       return 0;
129     }
130   }
131   if( pSet->n<N_OR_COST ){
132     p = &pSet->a[pSet->n++];
133     p->nOut = nOut;
134   }else{
135     p = pSet->a;
136     for(i=1; i<pSet->n; i++){
137       if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i;
138     }
139     if( p->rRun<=rRun ) return 0;
140   }
141 whereOrInsert_done:
142   p->prereq = prereq;
143   p->rRun = rRun;
144   if( p->nOut>nOut ) p->nOut = nOut;
145   return 1;
146 }
147 
148 /*
149 ** Return the bitmask for the given cursor number.  Return 0 if
150 ** iCursor is not in the set.
151 */
152 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){
153   int i;
154   assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 );
155   for(i=0; i<pMaskSet->n; i++){
156     if( pMaskSet->ix[i]==iCursor ){
157       return MASKBIT(i);
158     }
159   }
160   return 0;
161 }
162 
163 /*
164 ** Create a new mask for cursor iCursor.
165 **
166 ** There is one cursor per table in the FROM clause.  The number of
167 ** tables in the FROM clause is limited by a test early in the
168 ** sqlite3WhereBegin() routine.  So we know that the pMaskSet->ix[]
169 ** array will never overflow.
170 */
171 static void createMask(WhereMaskSet *pMaskSet, int iCursor){
172   assert( pMaskSet->n < ArraySize(pMaskSet->ix) );
173   pMaskSet->ix[pMaskSet->n++] = iCursor;
174 }
175 
176 /*
177 ** Advance to the next WhereTerm that matches according to the criteria
178 ** established when the pScan object was initialized by whereScanInit().
179 ** Return NULL if there are no more matching WhereTerms.
180 */
181 static WhereTerm *whereScanNext(WhereScan *pScan){
182   int iCur;            /* The cursor on the LHS of the term */
183   i16 iColumn;         /* The column on the LHS of the term.  -1 for IPK */
184   Expr *pX;            /* An expression being tested */
185   WhereClause *pWC;    /* Shorthand for pScan->pWC */
186   WhereTerm *pTerm;    /* The term being tested */
187   int k = pScan->k;    /* Where to start scanning */
188 
189   while( pScan->iEquiv<=pScan->nEquiv ){
190     iCur = pScan->aiCur[pScan->iEquiv-1];
191     iColumn = pScan->aiColumn[pScan->iEquiv-1];
192     if( iColumn==XN_EXPR && pScan->pIdxExpr==0 ) return 0;
193     while( (pWC = pScan->pWC)!=0 ){
194       for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){
195         if( pTerm->leftCursor==iCur
196          && pTerm->u.leftColumn==iColumn
197          && (iColumn!=XN_EXPR
198              || sqlite3ExprCompare(pTerm->pExpr->pLeft,pScan->pIdxExpr,iCur)==0)
199          && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin))
200         ){
201           if( (pTerm->eOperator & WO_EQUIV)!=0
202            && pScan->nEquiv<ArraySize(pScan->aiCur)
203            && (pX = sqlite3ExprSkipCollate(pTerm->pExpr->pRight))->op==TK_COLUMN
204           ){
205             int j;
206             for(j=0; j<pScan->nEquiv; j++){
207               if( pScan->aiCur[j]==pX->iTable
208                && pScan->aiColumn[j]==pX->iColumn ){
209                   break;
210               }
211             }
212             if( j==pScan->nEquiv ){
213               pScan->aiCur[j] = pX->iTable;
214               pScan->aiColumn[j] = pX->iColumn;
215               pScan->nEquiv++;
216             }
217           }
218           if( (pTerm->eOperator & pScan->opMask)!=0 ){
219             /* Verify the affinity and collating sequence match */
220             if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){
221               CollSeq *pColl;
222               Parse *pParse = pWC->pWInfo->pParse;
223               pX = pTerm->pExpr;
224               if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){
225                 continue;
226               }
227               assert(pX->pLeft);
228               pColl = sqlite3BinaryCompareCollSeq(pParse,
229                                                   pX->pLeft, pX->pRight);
230               if( pColl==0 ) pColl = pParse->db->pDfltColl;
231               if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){
232                 continue;
233               }
234             }
235             if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0
236              && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN
237              && pX->iTable==pScan->aiCur[0]
238              && pX->iColumn==pScan->aiColumn[0]
239             ){
240               testcase( pTerm->eOperator & WO_IS );
241               continue;
242             }
243             pScan->k = k+1;
244             return pTerm;
245           }
246         }
247       }
248       pScan->pWC = pScan->pWC->pOuter;
249       k = 0;
250     }
251     pScan->pWC = pScan->pOrigWC;
252     k = 0;
253     pScan->iEquiv++;
254   }
255   return 0;
256 }
257 
258 /*
259 ** Initialize a WHERE clause scanner object.  Return a pointer to the
260 ** first match.  Return NULL if there are no matches.
261 **
262 ** The scanner will be searching the WHERE clause pWC.  It will look
263 ** for terms of the form "X <op> <expr>" where X is column iColumn of table
264 ** iCur.  The <op> must be one of the operators described by opMask.
265 **
266 ** If the search is for X and the WHERE clause contains terms of the
267 ** form X=Y then this routine might also return terms of the form
268 ** "Y <op> <expr>".  The number of levels of transitivity is limited,
269 ** but is enough to handle most commonly occurring SQL statements.
270 **
271 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with
272 ** index pIdx.
273 */
274 static WhereTerm *whereScanInit(
275   WhereScan *pScan,       /* The WhereScan object being initialized */
276   WhereClause *pWC,       /* The WHERE clause to be scanned */
277   int iCur,               /* Cursor to scan for */
278   int iColumn,            /* Column to scan for */
279   u32 opMask,             /* Operator(s) to scan for */
280   Index *pIdx             /* Must be compatible with this index */
281 ){
282   int j = 0;
283 
284   /* memset(pScan, 0, sizeof(*pScan)); */
285   pScan->pOrigWC = pWC;
286   pScan->pWC = pWC;
287   pScan->pIdxExpr = 0;
288   if( pIdx ){
289     j = iColumn;
290     iColumn = pIdx->aiColumn[j];
291     if( iColumn==XN_EXPR ) pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr;
292     if( iColumn==pIdx->pTable->iPKey ) iColumn = XN_ROWID;
293   }
294   if( pIdx && iColumn>=0 ){
295     pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity;
296     pScan->zCollName = pIdx->azColl[j];
297   }else{
298     pScan->idxaff = 0;
299     pScan->zCollName = 0;
300   }
301   pScan->opMask = opMask;
302   pScan->k = 0;
303   pScan->aiCur[0] = iCur;
304   pScan->aiColumn[0] = iColumn;
305   pScan->nEquiv = 1;
306   pScan->iEquiv = 1;
307   return whereScanNext(pScan);
308 }
309 
310 /*
311 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>"
312 ** where X is a reference to the iColumn of table iCur and <op> is one of
313 ** the WO_xx operator codes specified by the op parameter.
314 ** Return a pointer to the term.  Return 0 if not found.
315 **
316 ** If pIdx!=0 then search for terms matching the iColumn-th column of pIdx
317 ** rather than the iColumn-th column of table iCur.
318 **
319 ** The term returned might by Y=<expr> if there is another constraint in
320 ** the WHERE clause that specifies that X=Y.  Any such constraints will be
321 ** identified by the WO_EQUIV bit in the pTerm->eOperator field.  The
322 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11
323 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10
324 ** other equivalent values.  Hence a search for X will return <expr> if X=A1
325 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>.
326 **
327 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>"
328 ** then try for the one with no dependencies on <expr> - in other words where
329 ** <expr> is a constant expression of some kind.  Only return entries of
330 ** the form "X <op> Y" where Y is a column in another table if no terms of
331 ** the form "X <op> <const-expr>" exist.   If no terms with a constant RHS
332 ** exist, try to return a term that does not use WO_EQUIV.
333 */
334 WhereTerm *sqlite3WhereFindTerm(
335   WhereClause *pWC,     /* The WHERE clause to be searched */
336   int iCur,             /* Cursor number of LHS */
337   int iColumn,          /* Column number of LHS */
338   Bitmask notReady,     /* RHS must not overlap with this mask */
339   u32 op,               /* Mask of WO_xx values describing operator */
340   Index *pIdx           /* Must be compatible with this index, if not NULL */
341 ){
342   WhereTerm *pResult = 0;
343   WhereTerm *p;
344   WhereScan scan;
345 
346   p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx);
347   op &= WO_EQ|WO_IS;
348   while( p ){
349     if( (p->prereqRight & notReady)==0 ){
350       if( p->prereqRight==0 && (p->eOperator&op)!=0 ){
351         testcase( p->eOperator & WO_IS );
352         return p;
353       }
354       if( pResult==0 ) pResult = p;
355     }
356     p = whereScanNext(&scan);
357   }
358   return pResult;
359 }
360 
361 /*
362 ** This function searches pList for an entry that matches the iCol-th column
363 ** of index pIdx.
364 **
365 ** If such an expression is found, its index in pList->a[] is returned. If
366 ** no expression is found, -1 is returned.
367 */
368 static int findIndexCol(
369   Parse *pParse,                  /* Parse context */
370   ExprList *pList,                /* Expression list to search */
371   int iBase,                      /* Cursor for table associated with pIdx */
372   Index *pIdx,                    /* Index to match column of */
373   int iCol                        /* Column of index to match */
374 ){
375   int i;
376   const char *zColl = pIdx->azColl[iCol];
377 
378   for(i=0; i<pList->nExpr; i++){
379     Expr *p = sqlite3ExprSkipCollate(pList->a[i].pExpr);
380     if( p->op==TK_COLUMN
381      && p->iColumn==pIdx->aiColumn[iCol]
382      && p->iTable==iBase
383     ){
384       CollSeq *pColl = sqlite3ExprCollSeq(pParse, pList->a[i].pExpr);
385       if( pColl && 0==sqlite3StrICmp(pColl->zName, zColl) ){
386         return i;
387       }
388     }
389   }
390 
391   return -1;
392 }
393 
394 /*
395 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL
396 */
397 static int indexColumnNotNull(Index *pIdx, int iCol){
398   int j;
399   assert( pIdx!=0 );
400   assert( iCol>=0 && iCol<pIdx->nColumn );
401   j = pIdx->aiColumn[iCol];
402   if( j>=0 ){
403     return pIdx->pTable->aCol[j].notNull;
404   }else if( j==(-1) ){
405     return 1;
406   }else{
407     assert( j==(-2) );
408     return 0;  /* Assume an indexed expression can always yield a NULL */
409 
410   }
411 }
412 
413 /*
414 ** Return true if the DISTINCT expression-list passed as the third argument
415 ** is redundant.
416 **
417 ** A DISTINCT list is redundant if any subset of the columns in the
418 ** DISTINCT list are collectively unique and individually non-null.
419 */
420 static int isDistinctRedundant(
421   Parse *pParse,            /* Parsing context */
422   SrcList *pTabList,        /* The FROM clause */
423   WhereClause *pWC,         /* The WHERE clause */
424   ExprList *pDistinct       /* The result set that needs to be DISTINCT */
425 ){
426   Table *pTab;
427   Index *pIdx;
428   int i;
429   int iBase;
430 
431   /* If there is more than one table or sub-select in the FROM clause of
432   ** this query, then it will not be possible to show that the DISTINCT
433   ** clause is redundant. */
434   if( pTabList->nSrc!=1 ) return 0;
435   iBase = pTabList->a[0].iCursor;
436   pTab = pTabList->a[0].pTab;
437 
438   /* If any of the expressions is an IPK column on table iBase, then return
439   ** true. Note: The (p->iTable==iBase) part of this test may be false if the
440   ** current SELECT is a correlated sub-query.
441   */
442   for(i=0; i<pDistinct->nExpr; i++){
443     Expr *p = sqlite3ExprSkipCollate(pDistinct->a[i].pExpr);
444     if( p->op==TK_COLUMN && p->iTable==iBase && p->iColumn<0 ) return 1;
445   }
446 
447   /* Loop through all indices on the table, checking each to see if it makes
448   ** the DISTINCT qualifier redundant. It does so if:
449   **
450   **   1. The index is itself UNIQUE, and
451   **
452   **   2. All of the columns in the index are either part of the pDistinct
453   **      list, or else the WHERE clause contains a term of the form "col=X",
454   **      where X is a constant value. The collation sequences of the
455   **      comparison and select-list expressions must match those of the index.
456   **
457   **   3. All of those index columns for which the WHERE clause does not
458   **      contain a "col=X" term are subject to a NOT NULL constraint.
459   */
460   for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
461     if( !IsUniqueIndex(pIdx) ) continue;
462     for(i=0; i<pIdx->nKeyCol; i++){
463       if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){
464         if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break;
465         if( indexColumnNotNull(pIdx, i)==0 ) break;
466       }
467     }
468     if( i==pIdx->nKeyCol ){
469       /* This index implies that the DISTINCT qualifier is redundant. */
470       return 1;
471     }
472   }
473 
474   return 0;
475 }
476 
477 
478 /*
479 ** Estimate the logarithm of the input value to base 2.
480 */
481 static LogEst estLog(LogEst N){
482   return N<=10 ? 0 : sqlite3LogEst(N) - 33;
483 }
484 
485 /*
486 ** Convert OP_Column opcodes to OP_Copy in previously generated code.
487 **
488 ** This routine runs over generated VDBE code and translates OP_Column
489 ** opcodes into OP_Copy when the table is being accessed via co-routine
490 ** instead of via table lookup.
491 **
492 ** If the bIncrRowid parameter is 0, then any OP_Rowid instructions on
493 ** cursor iTabCur are transformed into OP_Null. Or, if bIncrRowid is non-zero,
494 ** then each OP_Rowid is transformed into an instruction to increment the
495 ** value stored in its output register.
496 */
497 static void translateColumnToCopy(
498   Vdbe *v,            /* The VDBE containing code to translate */
499   int iStart,         /* Translate from this opcode to the end */
500   int iTabCur,        /* OP_Column/OP_Rowid references to this table */
501   int iRegister,      /* The first column is in this register */
502   int bIncrRowid      /* If non-zero, transform OP_rowid to OP_AddImm(1) */
503 ){
504   VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart);
505   int iEnd = sqlite3VdbeCurrentAddr(v);
506   for(; iStart<iEnd; iStart++, pOp++){
507     if( pOp->p1!=iTabCur ) continue;
508     if( pOp->opcode==OP_Column ){
509       pOp->opcode = OP_Copy;
510       pOp->p1 = pOp->p2 + iRegister;
511       pOp->p2 = pOp->p3;
512       pOp->p3 = 0;
513     }else if( pOp->opcode==OP_Rowid ){
514       if( bIncrRowid ){
515         /* Increment the value stored in the P2 operand of the OP_Rowid. */
516         pOp->opcode = OP_AddImm;
517         pOp->p1 = pOp->p2;
518         pOp->p2 = 1;
519       }else{
520         pOp->opcode = OP_Null;
521         pOp->p1 = 0;
522         pOp->p3 = 0;
523       }
524     }
525   }
526 }
527 
528 /*
529 ** Two routines for printing the content of an sqlite3_index_info
530 ** structure.  Used for testing and debugging only.  If neither
531 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines
532 ** are no-ops.
533 */
534 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED)
535 static void TRACE_IDX_INPUTS(sqlite3_index_info *p){
536   int i;
537   if( !sqlite3WhereTrace ) return;
538   for(i=0; i<p->nConstraint; i++){
539     sqlite3DebugPrintf("  constraint[%d]: col=%d termid=%d op=%d usabled=%d\n",
540        i,
541        p->aConstraint[i].iColumn,
542        p->aConstraint[i].iTermOffset,
543        p->aConstraint[i].op,
544        p->aConstraint[i].usable);
545   }
546   for(i=0; i<p->nOrderBy; i++){
547     sqlite3DebugPrintf("  orderby[%d]: col=%d desc=%d\n",
548        i,
549        p->aOrderBy[i].iColumn,
550        p->aOrderBy[i].desc);
551   }
552 }
553 static void TRACE_IDX_OUTPUTS(sqlite3_index_info *p){
554   int i;
555   if( !sqlite3WhereTrace ) return;
556   for(i=0; i<p->nConstraint; i++){
557     sqlite3DebugPrintf("  usage[%d]: argvIdx=%d omit=%d\n",
558        i,
559        p->aConstraintUsage[i].argvIndex,
560        p->aConstraintUsage[i].omit);
561   }
562   sqlite3DebugPrintf("  idxNum=%d\n", p->idxNum);
563   sqlite3DebugPrintf("  idxStr=%s\n", p->idxStr);
564   sqlite3DebugPrintf("  orderByConsumed=%d\n", p->orderByConsumed);
565   sqlite3DebugPrintf("  estimatedCost=%g\n", p->estimatedCost);
566   sqlite3DebugPrintf("  estimatedRows=%lld\n", p->estimatedRows);
567 }
568 #else
569 #define TRACE_IDX_INPUTS(A)
570 #define TRACE_IDX_OUTPUTS(A)
571 #endif
572 
573 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
574 /*
575 ** Return TRUE if the WHERE clause term pTerm is of a form where it
576 ** could be used with an index to access pSrc, assuming an appropriate
577 ** index existed.
578 */
579 static int termCanDriveIndex(
580   WhereTerm *pTerm,              /* WHERE clause term to check */
581   struct SrcList_item *pSrc,     /* Table we are trying to access */
582   Bitmask notReady               /* Tables in outer loops of the join */
583 ){
584   char aff;
585   if( pTerm->leftCursor!=pSrc->iCursor ) return 0;
586   if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0;
587   if( (pTerm->prereqRight & notReady)!=0 ) return 0;
588   if( pTerm->u.leftColumn<0 ) return 0;
589   aff = pSrc->pTab->aCol[pTerm->u.leftColumn].affinity;
590   if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0;
591   testcase( pTerm->pExpr->op==TK_IS );
592   return 1;
593 }
594 #endif
595 
596 
597 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
598 /*
599 ** Generate code to construct the Index object for an automatic index
600 ** and to set up the WhereLevel object pLevel so that the code generator
601 ** makes use of the automatic index.
602 */
603 static void constructAutomaticIndex(
604   Parse *pParse,              /* The parsing context */
605   WhereClause *pWC,           /* The WHERE clause */
606   struct SrcList_item *pSrc,  /* The FROM clause term to get the next index */
607   Bitmask notReady,           /* Mask of cursors that are not available */
608   WhereLevel *pLevel          /* Write new index here */
609 ){
610   int nKeyCol;                /* Number of columns in the constructed index */
611   WhereTerm *pTerm;           /* A single term of the WHERE clause */
612   WhereTerm *pWCEnd;          /* End of pWC->a[] */
613   Index *pIdx;                /* Object describing the transient index */
614   Vdbe *v;                    /* Prepared statement under construction */
615   int addrInit;               /* Address of the initialization bypass jump */
616   Table *pTable;              /* The table being indexed */
617   int addrTop;                /* Top of the index fill loop */
618   int regRecord;              /* Register holding an index record */
619   int n;                      /* Column counter */
620   int i;                      /* Loop counter */
621   int mxBitCol;               /* Maximum column in pSrc->colUsed */
622   CollSeq *pColl;             /* Collating sequence to on a column */
623   WhereLoop *pLoop;           /* The Loop object */
624   char *zNotUsed;             /* Extra space on the end of pIdx */
625   Bitmask idxCols;            /* Bitmap of columns used for indexing */
626   Bitmask extraCols;          /* Bitmap of additional columns */
627   u8 sentWarning = 0;         /* True if a warnning has been issued */
628   Expr *pPartial = 0;         /* Partial Index Expression */
629   int iContinue = 0;          /* Jump here to skip excluded rows */
630   struct SrcList_item *pTabItem;  /* FROM clause term being indexed */
631   int addrCounter = 0;        /* Address where integer counter is initialized */
632   int regBase;                /* Array of registers where record is assembled */
633 
634   /* Generate code to skip over the creation and initialization of the
635   ** transient index on 2nd and subsequent iterations of the loop. */
636   v = pParse->pVdbe;
637   assert( v!=0 );
638   addrInit = sqlite3CodeOnce(pParse); VdbeCoverage(v);
639 
640   /* Count the number of columns that will be added to the index
641   ** and used to match WHERE clause constraints */
642   nKeyCol = 0;
643   pTable = pSrc->pTab;
644   pWCEnd = &pWC->a[pWC->nTerm];
645   pLoop = pLevel->pWLoop;
646   idxCols = 0;
647   for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
648     Expr *pExpr = pTerm->pExpr;
649     assert( !ExprHasProperty(pExpr, EP_FromJoin)    /* prereq always non-zero */
650          || pExpr->iRightJoinTable!=pSrc->iCursor   /*   for the right-hand   */
651          || pLoop->prereq!=0 );                     /*   table of a LEFT JOIN */
652     if( pLoop->prereq==0
653      && (pTerm->wtFlags & TERM_VIRTUAL)==0
654      && !ExprHasProperty(pExpr, EP_FromJoin)
655      && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){
656       pPartial = sqlite3ExprAnd(pParse->db, pPartial,
657                                 sqlite3ExprDup(pParse->db, pExpr, 0));
658     }
659     if( termCanDriveIndex(pTerm, pSrc, notReady) ){
660       int iCol = pTerm->u.leftColumn;
661       Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol);
662       testcase( iCol==BMS );
663       testcase( iCol==BMS-1 );
664       if( !sentWarning ){
665         sqlite3_log(SQLITE_WARNING_AUTOINDEX,
666             "automatic index on %s(%s)", pTable->zName,
667             pTable->aCol[iCol].zName);
668         sentWarning = 1;
669       }
670       if( (idxCols & cMask)==0 ){
671         if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){
672           goto end_auto_index_create;
673         }
674         pLoop->aLTerm[nKeyCol++] = pTerm;
675         idxCols |= cMask;
676       }
677     }
678   }
679   assert( nKeyCol>0 );
680   pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol;
681   pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED
682                      | WHERE_AUTO_INDEX;
683 
684   /* Count the number of additional columns needed to create a
685   ** covering index.  A "covering index" is an index that contains all
686   ** columns that are needed by the query.  With a covering index, the
687   ** original table never needs to be accessed.  Automatic indices must
688   ** be a covering index because the index will not be updated if the
689   ** original table changes and the index and table cannot both be used
690   ** if they go out of sync.
691   */
692   extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1));
693   mxBitCol = MIN(BMS-1,pTable->nCol);
694   testcase( pTable->nCol==BMS-1 );
695   testcase( pTable->nCol==BMS-2 );
696   for(i=0; i<mxBitCol; i++){
697     if( extraCols & MASKBIT(i) ) nKeyCol++;
698   }
699   if( pSrc->colUsed & MASKBIT(BMS-1) ){
700     nKeyCol += pTable->nCol - BMS + 1;
701   }
702 
703   /* Construct the Index object to describe this index */
704   pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed);
705   if( pIdx==0 ) goto end_auto_index_create;
706   pLoop->u.btree.pIndex = pIdx;
707   pIdx->zName = "auto-index";
708   pIdx->pTable = pTable;
709   n = 0;
710   idxCols = 0;
711   for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){
712     if( termCanDriveIndex(pTerm, pSrc, notReady) ){
713       int iCol = pTerm->u.leftColumn;
714       Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol);
715       testcase( iCol==BMS-1 );
716       testcase( iCol==BMS );
717       if( (idxCols & cMask)==0 ){
718         Expr *pX = pTerm->pExpr;
719         idxCols |= cMask;
720         pIdx->aiColumn[n] = pTerm->u.leftColumn;
721         pColl = sqlite3BinaryCompareCollSeq(pParse, pX->pLeft, pX->pRight);
722         pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY;
723         n++;
724       }
725     }
726   }
727   assert( (u32)n==pLoop->u.btree.nEq );
728 
729   /* Add additional columns needed to make the automatic index into
730   ** a covering index */
731   for(i=0; i<mxBitCol; i++){
732     if( extraCols & MASKBIT(i) ){
733       pIdx->aiColumn[n] = i;
734       pIdx->azColl[n] = sqlite3StrBINARY;
735       n++;
736     }
737   }
738   if( pSrc->colUsed & MASKBIT(BMS-1) ){
739     for(i=BMS-1; i<pTable->nCol; i++){
740       pIdx->aiColumn[n] = i;
741       pIdx->azColl[n] = sqlite3StrBINARY;
742       n++;
743     }
744   }
745   assert( n==nKeyCol );
746   pIdx->aiColumn[n] = XN_ROWID;
747   pIdx->azColl[n] = sqlite3StrBINARY;
748 
749   /* Create the automatic index */
750   assert( pLevel->iIdxCur>=0 );
751   pLevel->iIdxCur = pParse->nTab++;
752   sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1);
753   sqlite3VdbeSetP4KeyInfo(pParse, pIdx);
754   VdbeComment((v, "for %s", pTable->zName));
755 
756   /* Fill the automatic index with content */
757   sqlite3ExprCachePush(pParse);
758   pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom];
759   if( pTabItem->fg.viaCoroutine ){
760     int regYield = pTabItem->regReturn;
761     addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0);
762     sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub);
763     addrTop =  sqlite3VdbeAddOp1(v, OP_Yield, regYield);
764     VdbeCoverage(v);
765     VdbeComment((v, "next row of \"%s\"", pTabItem->pTab->zName));
766   }else{
767     addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v);
768   }
769   if( pPartial ){
770     iContinue = sqlite3VdbeMakeLabel(v);
771     sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL);
772     pLoop->wsFlags |= WHERE_PARTIALIDX;
773   }
774   regRecord = sqlite3GetTempReg(pParse);
775   regBase = sqlite3GenerateIndexKey(
776       pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0
777   );
778   sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord);
779   sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT);
780   if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue);
781   if( pTabItem->fg.viaCoroutine ){
782     sqlite3VdbeChangeP2(v, addrCounter, regBase+n);
783     translateColumnToCopy(v, addrTop, pLevel->iTabCur, pTabItem->regResult, 1);
784     sqlite3VdbeGoto(v, addrTop);
785     pTabItem->fg.viaCoroutine = 0;
786   }else{
787     sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v);
788   }
789   sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX);
790   sqlite3VdbeJumpHere(v, addrTop);
791   sqlite3ReleaseTempReg(pParse, regRecord);
792   sqlite3ExprCachePop(pParse);
793 
794   /* Jump here when skipping the initialization */
795   sqlite3VdbeJumpHere(v, addrInit);
796 
797 end_auto_index_create:
798   sqlite3ExprDelete(pParse->db, pPartial);
799 }
800 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
801 
802 #ifndef SQLITE_OMIT_VIRTUALTABLE
803 /*
804 ** Allocate and populate an sqlite3_index_info structure. It is the
805 ** responsibility of the caller to eventually release the structure
806 ** by passing the pointer returned by this function to sqlite3_free().
807 */
808 static sqlite3_index_info *allocateIndexInfo(
809   Parse *pParse,
810   WhereClause *pWC,
811   Bitmask mUnusable,              /* Ignore terms with these prereqs */
812   struct SrcList_item *pSrc,
813   ExprList *pOrderBy
814 ){
815   int i, j;
816   int nTerm;
817   struct sqlite3_index_constraint *pIdxCons;
818   struct sqlite3_index_orderby *pIdxOrderBy;
819   struct sqlite3_index_constraint_usage *pUsage;
820   WhereTerm *pTerm;
821   int nOrderBy;
822   sqlite3_index_info *pIdxInfo;
823 
824   /* Count the number of possible WHERE clause constraints referring
825   ** to this virtual table */
826   for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
827     if( pTerm->leftCursor != pSrc->iCursor ) continue;
828     if( pTerm->prereqRight & mUnusable ) continue;
829     assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) );
830     testcase( pTerm->eOperator & WO_IN );
831     testcase( pTerm->eOperator & WO_ISNULL );
832     testcase( pTerm->eOperator & WO_IS );
833     testcase( pTerm->eOperator & WO_ALL );
834     if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue;
835     if( pTerm->wtFlags & TERM_VNULL ) continue;
836     assert( pTerm->u.leftColumn>=(-1) );
837     nTerm++;
838   }
839 
840   /* If the ORDER BY clause contains only columns in the current
841   ** virtual table then allocate space for the aOrderBy part of
842   ** the sqlite3_index_info structure.
843   */
844   nOrderBy = 0;
845   if( pOrderBy ){
846     int n = pOrderBy->nExpr;
847     for(i=0; i<n; i++){
848       Expr *pExpr = pOrderBy->a[i].pExpr;
849       if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break;
850     }
851     if( i==n){
852       nOrderBy = n;
853     }
854   }
855 
856   /* Allocate the sqlite3_index_info structure
857   */
858   pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo)
859                            + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm
860                            + sizeof(*pIdxOrderBy)*nOrderBy );
861   if( pIdxInfo==0 ){
862     sqlite3ErrorMsg(pParse, "out of memory");
863     return 0;
864   }
865 
866   /* Initialize the structure.  The sqlite3_index_info structure contains
867   ** many fields that are declared "const" to prevent xBestIndex from
868   ** changing them.  We have to do some funky casting in order to
869   ** initialize those fields.
870   */
871   pIdxCons = (struct sqlite3_index_constraint*)&pIdxInfo[1];
872   pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm];
873   pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy];
874   *(int*)&pIdxInfo->nConstraint = nTerm;
875   *(int*)&pIdxInfo->nOrderBy = nOrderBy;
876   *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint = pIdxCons;
877   *(struct sqlite3_index_orderby**)&pIdxInfo->aOrderBy = pIdxOrderBy;
878   *(struct sqlite3_index_constraint_usage**)&pIdxInfo->aConstraintUsage =
879                                                                    pUsage;
880 
881   for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
882     u8 op;
883     if( pTerm->leftCursor != pSrc->iCursor ) continue;
884     if( pTerm->prereqRight & mUnusable ) continue;
885     assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) );
886     testcase( pTerm->eOperator & WO_IN );
887     testcase( pTerm->eOperator & WO_IS );
888     testcase( pTerm->eOperator & WO_ISNULL );
889     testcase( pTerm->eOperator & WO_ALL );
890     if( (pTerm->eOperator & ~(WO_ISNULL|WO_EQUIV|WO_IS))==0 ) continue;
891     if( pTerm->wtFlags & TERM_VNULL ) continue;
892     assert( pTerm->u.leftColumn>=(-1) );
893     pIdxCons[j].iColumn = pTerm->u.leftColumn;
894     pIdxCons[j].iTermOffset = i;
895     op = (u8)pTerm->eOperator & WO_ALL;
896     if( op==WO_IN ) op = WO_EQ;
897     if( op==WO_MATCH ){
898       op = pTerm->eMatchOp;
899     }
900     pIdxCons[j].op = op;
901     /* The direct assignment in the previous line is possible only because
902     ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical.  The
903     ** following asserts verify this fact. */
904     assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ );
905     assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT );
906     assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE );
907     assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT );
908     assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE );
909     assert( WO_MATCH==SQLITE_INDEX_CONSTRAINT_MATCH );
910     assert( pTerm->eOperator & (WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_MATCH) );
911     j++;
912   }
913   for(i=0; i<nOrderBy; i++){
914     Expr *pExpr = pOrderBy->a[i].pExpr;
915     pIdxOrderBy[i].iColumn = pExpr->iColumn;
916     pIdxOrderBy[i].desc = pOrderBy->a[i].sortOrder;
917   }
918 
919   return pIdxInfo;
920 }
921 
922 /*
923 ** The table object reference passed as the second argument to this function
924 ** must represent a virtual table. This function invokes the xBestIndex()
925 ** method of the virtual table with the sqlite3_index_info object that
926 ** comes in as the 3rd argument to this function.
927 **
928 ** If an error occurs, pParse is populated with an error message and a
929 ** non-zero value is returned. Otherwise, 0 is returned and the output
930 ** part of the sqlite3_index_info structure is left populated.
931 **
932 ** Whether or not an error is returned, it is the responsibility of the
933 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates
934 ** that this is required.
935 */
936 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){
937   sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab;
938   int rc;
939 
940   TRACE_IDX_INPUTS(p);
941   rc = pVtab->pModule->xBestIndex(pVtab, p);
942   TRACE_IDX_OUTPUTS(p);
943 
944   if( rc!=SQLITE_OK ){
945     if( rc==SQLITE_NOMEM ){
946       sqlite3OomFault(pParse->db);
947     }else if( !pVtab->zErrMsg ){
948       sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc));
949     }else{
950       sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg);
951     }
952   }
953   sqlite3_free(pVtab->zErrMsg);
954   pVtab->zErrMsg = 0;
955 
956 #if 0
957   /* This error is now caught by the caller.
958   ** Search for "xBestIndex malfunction" below */
959   for(i=0; i<p->nConstraint; i++){
960     if( !p->aConstraint[i].usable && p->aConstraintUsage[i].argvIndex>0 ){
961       sqlite3ErrorMsg(pParse,
962           "table %s: xBestIndex returned an invalid plan", pTab->zName);
963     }
964   }
965 #endif
966 
967   return pParse->nErr;
968 }
969 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */
970 
971 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
972 /*
973 ** Estimate the location of a particular key among all keys in an
974 ** index.  Store the results in aStat as follows:
975 **
976 **    aStat[0]      Est. number of rows less than pRec
977 **    aStat[1]      Est. number of rows equal to pRec
978 **
979 ** Return the index of the sample that is the smallest sample that
980 ** is greater than or equal to pRec. Note that this index is not an index
981 ** into the aSample[] array - it is an index into a virtual set of samples
982 ** based on the contents of aSample[] and the number of fields in record
983 ** pRec.
984 */
985 static int whereKeyStats(
986   Parse *pParse,              /* Database connection */
987   Index *pIdx,                /* Index to consider domain of */
988   UnpackedRecord *pRec,       /* Vector of values to consider */
989   int roundUp,                /* Round up if true.  Round down if false */
990   tRowcnt *aStat              /* OUT: stats written here */
991 ){
992   IndexSample *aSample = pIdx->aSample;
993   int iCol;                   /* Index of required stats in anEq[] etc. */
994   int i;                      /* Index of first sample >= pRec */
995   int iSample;                /* Smallest sample larger than or equal to pRec */
996   int iMin = 0;               /* Smallest sample not yet tested */
997   int iTest;                  /* Next sample to test */
998   int res;                    /* Result of comparison operation */
999   int nField;                 /* Number of fields in pRec */
1000   tRowcnt iLower = 0;         /* anLt[] + anEq[] of largest sample pRec is > */
1001 
1002 #ifndef SQLITE_DEBUG
1003   UNUSED_PARAMETER( pParse );
1004 #endif
1005   assert( pRec!=0 );
1006   assert( pIdx->nSample>0 );
1007   assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol );
1008 
1009   /* Do a binary search to find the first sample greater than or equal
1010   ** to pRec. If pRec contains a single field, the set of samples to search
1011   ** is simply the aSample[] array. If the samples in aSample[] contain more
1012   ** than one fields, all fields following the first are ignored.
1013   **
1014   ** If pRec contains N fields, where N is more than one, then as well as the
1015   ** samples in aSample[] (truncated to N fields), the search also has to
1016   ** consider prefixes of those samples. For example, if the set of samples
1017   ** in aSample is:
1018   **
1019   **     aSample[0] = (a, 5)
1020   **     aSample[1] = (a, 10)
1021   **     aSample[2] = (b, 5)
1022   **     aSample[3] = (c, 100)
1023   **     aSample[4] = (c, 105)
1024   **
1025   ** Then the search space should ideally be the samples above and the
1026   ** unique prefixes [a], [b] and [c]. But since that is hard to organize,
1027   ** the code actually searches this set:
1028   **
1029   **     0: (a)
1030   **     1: (a, 5)
1031   **     2: (a, 10)
1032   **     3: (a, 10)
1033   **     4: (b)
1034   **     5: (b, 5)
1035   **     6: (c)
1036   **     7: (c, 100)
1037   **     8: (c, 105)
1038   **     9: (c, 105)
1039   **
1040   ** For each sample in the aSample[] array, N samples are present in the
1041   ** effective sample array. In the above, samples 0 and 1 are based on
1042   ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc.
1043   **
1044   ** Often, sample i of each block of N effective samples has (i+1) fields.
1045   ** Except, each sample may be extended to ensure that it is greater than or
1046   ** equal to the previous sample in the array. For example, in the above,
1047   ** sample 2 is the first sample of a block of N samples, so at first it
1048   ** appears that it should be 1 field in size. However, that would make it
1049   ** smaller than sample 1, so the binary search would not work. As a result,
1050   ** it is extended to two fields. The duplicates that this creates do not
1051   ** cause any problems.
1052   */
1053   nField = pRec->nField;
1054   iCol = 0;
1055   iSample = pIdx->nSample * nField;
1056   do{
1057     int iSamp;                    /* Index in aSample[] of test sample */
1058     int n;                        /* Number of fields in test sample */
1059 
1060     iTest = (iMin+iSample)/2;
1061     iSamp = iTest / nField;
1062     if( iSamp>0 ){
1063       /* The proposed effective sample is a prefix of sample aSample[iSamp].
1064       ** Specifically, the shortest prefix of at least (1 + iTest%nField)
1065       ** fields that is greater than the previous effective sample.  */
1066       for(n=(iTest % nField) + 1; n<nField; n++){
1067         if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break;
1068       }
1069     }else{
1070       n = iTest + 1;
1071     }
1072 
1073     pRec->nField = n;
1074     res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec);
1075     if( res<0 ){
1076       iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1];
1077       iMin = iTest+1;
1078     }else if( res==0 && n<nField ){
1079       iLower = aSample[iSamp].anLt[n-1];
1080       iMin = iTest+1;
1081       res = -1;
1082     }else{
1083       iSample = iTest;
1084       iCol = n-1;
1085     }
1086   }while( res && iMin<iSample );
1087   i = iSample / nField;
1088 
1089 #ifdef SQLITE_DEBUG
1090   /* The following assert statements check that the binary search code
1091   ** above found the right answer. This block serves no purpose other
1092   ** than to invoke the asserts.  */
1093   if( pParse->db->mallocFailed==0 ){
1094     if( res==0 ){
1095       /* If (res==0) is true, then pRec must be equal to sample i. */
1096       assert( i<pIdx->nSample );
1097       assert( iCol==nField-1 );
1098       pRec->nField = nField;
1099       assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)
1100            || pParse->db->mallocFailed
1101       );
1102     }else{
1103       /* Unless i==pIdx->nSample, indicating that pRec is larger than
1104       ** all samples in the aSample[] array, pRec must be smaller than the
1105       ** (iCol+1) field prefix of sample i.  */
1106       assert( i<=pIdx->nSample && i>=0 );
1107       pRec->nField = iCol+1;
1108       assert( i==pIdx->nSample
1109            || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0
1110            || pParse->db->mallocFailed );
1111 
1112       /* if i==0 and iCol==0, then record pRec is smaller than all samples
1113       ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must
1114       ** be greater than or equal to the (iCol) field prefix of sample i.
1115       ** If (i>0), then pRec must also be greater than sample (i-1).  */
1116       if( iCol>0 ){
1117         pRec->nField = iCol;
1118         assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0
1119              || pParse->db->mallocFailed );
1120       }
1121       if( i>0 ){
1122         pRec->nField = nField;
1123         assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0
1124              || pParse->db->mallocFailed );
1125       }
1126     }
1127   }
1128 #endif /* ifdef SQLITE_DEBUG */
1129 
1130   if( res==0 ){
1131     /* Record pRec is equal to sample i */
1132     assert( iCol==nField-1 );
1133     aStat[0] = aSample[i].anLt[iCol];
1134     aStat[1] = aSample[i].anEq[iCol];
1135   }else{
1136     /* At this point, the (iCol+1) field prefix of aSample[i] is the first
1137     ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec
1138     ** is larger than all samples in the array. */
1139     tRowcnt iUpper, iGap;
1140     if( i>=pIdx->nSample ){
1141       iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]);
1142     }else{
1143       iUpper = aSample[i].anLt[iCol];
1144     }
1145 
1146     if( iLower>=iUpper ){
1147       iGap = 0;
1148     }else{
1149       iGap = iUpper - iLower;
1150     }
1151     if( roundUp ){
1152       iGap = (iGap*2)/3;
1153     }else{
1154       iGap = iGap/3;
1155     }
1156     aStat[0] = iLower + iGap;
1157     aStat[1] = pIdx->aAvgEq[iCol];
1158   }
1159 
1160   /* Restore the pRec->nField value before returning.  */
1161   pRec->nField = nField;
1162   return i;
1163 }
1164 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1165 
1166 /*
1167 ** If it is not NULL, pTerm is a term that provides an upper or lower
1168 ** bound on a range scan. Without considering pTerm, it is estimated
1169 ** that the scan will visit nNew rows. This function returns the number
1170 ** estimated to be visited after taking pTerm into account.
1171 **
1172 ** If the user explicitly specified a likelihood() value for this term,
1173 ** then the return value is the likelihood multiplied by the number of
1174 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term
1175 ** has a likelihood of 0.50, and any other term a likelihood of 0.25.
1176 */
1177 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){
1178   LogEst nRet = nNew;
1179   if( pTerm ){
1180     if( pTerm->truthProb<=0 ){
1181       nRet += pTerm->truthProb;
1182     }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){
1183       nRet -= 20;        assert( 20==sqlite3LogEst(4) );
1184     }
1185   }
1186   return nRet;
1187 }
1188 
1189 
1190 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1191 /*
1192 ** Return the affinity for a single column of an index.
1193 */
1194 static char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){
1195   assert( iCol>=0 && iCol<pIdx->nColumn );
1196   if( !pIdx->zColAff ){
1197     if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB;
1198   }
1199   return pIdx->zColAff[iCol];
1200 }
1201 #endif
1202 
1203 
1204 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1205 /*
1206 ** This function is called to estimate the number of rows visited by a
1207 ** range-scan on a skip-scan index. For example:
1208 **
1209 **   CREATE INDEX i1 ON t1(a, b, c);
1210 **   SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?;
1211 **
1212 ** Value pLoop->nOut is currently set to the estimated number of rows
1213 ** visited for scanning (a=? AND b=?). This function reduces that estimate
1214 ** by some factor to account for the (c BETWEEN ? AND ?) expression based
1215 ** on the stat4 data for the index. this scan will be peformed multiple
1216 ** times (once for each (a,b) combination that matches a=?) is dealt with
1217 ** by the caller.
1218 **
1219 ** It does this by scanning through all stat4 samples, comparing values
1220 ** extracted from pLower and pUpper with the corresponding column in each
1221 ** sample. If L and U are the number of samples found to be less than or
1222 ** equal to the values extracted from pLower and pUpper respectively, and
1223 ** N is the total number of samples, the pLoop->nOut value is adjusted
1224 ** as follows:
1225 **
1226 **   nOut = nOut * ( min(U - L, 1) / N )
1227 **
1228 ** If pLower is NULL, or a value cannot be extracted from the term, L is
1229 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it,
1230 ** U is set to N.
1231 **
1232 ** Normally, this function sets *pbDone to 1 before returning. However,
1233 ** if no value can be extracted from either pLower or pUpper (and so the
1234 ** estimate of the number of rows delivered remains unchanged), *pbDone
1235 ** is left as is.
1236 **
1237 ** If an error occurs, an SQLite error code is returned. Otherwise,
1238 ** SQLITE_OK.
1239 */
1240 static int whereRangeSkipScanEst(
1241   Parse *pParse,       /* Parsing & code generating context */
1242   WhereTerm *pLower,   /* Lower bound on the range. ex: "x>123" Might be NULL */
1243   WhereTerm *pUpper,   /* Upper bound on the range. ex: "x<455" Might be NULL */
1244   WhereLoop *pLoop,    /* Update the .nOut value of this loop */
1245   int *pbDone          /* Set to true if at least one expr. value extracted */
1246 ){
1247   Index *p = pLoop->u.btree.pIndex;
1248   int nEq = pLoop->u.btree.nEq;
1249   sqlite3 *db = pParse->db;
1250   int nLower = -1;
1251   int nUpper = p->nSample+1;
1252   int rc = SQLITE_OK;
1253   u8 aff = sqlite3IndexColumnAffinity(db, p, nEq);
1254   CollSeq *pColl;
1255 
1256   sqlite3_value *p1 = 0;          /* Value extracted from pLower */
1257   sqlite3_value *p2 = 0;          /* Value extracted from pUpper */
1258   sqlite3_value *pVal = 0;        /* Value extracted from record */
1259 
1260   pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]);
1261   if( pLower ){
1262     rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1);
1263     nLower = 0;
1264   }
1265   if( pUpper && rc==SQLITE_OK ){
1266     rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2);
1267     nUpper = p2 ? 0 : p->nSample;
1268   }
1269 
1270   if( p1 || p2 ){
1271     int i;
1272     int nDiff;
1273     for(i=0; rc==SQLITE_OK && i<p->nSample; i++){
1274       rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal);
1275       if( rc==SQLITE_OK && p1 ){
1276         int res = sqlite3MemCompare(p1, pVal, pColl);
1277         if( res>=0 ) nLower++;
1278       }
1279       if( rc==SQLITE_OK && p2 ){
1280         int res = sqlite3MemCompare(p2, pVal, pColl);
1281         if( res>=0 ) nUpper++;
1282       }
1283     }
1284     nDiff = (nUpper - nLower);
1285     if( nDiff<=0 ) nDiff = 1;
1286 
1287     /* If there is both an upper and lower bound specified, and the
1288     ** comparisons indicate that they are close together, use the fallback
1289     ** method (assume that the scan visits 1/64 of the rows) for estimating
1290     ** the number of rows visited. Otherwise, estimate the number of rows
1291     ** using the method described in the header comment for this function. */
1292     if( nDiff!=1 || pUpper==0 || pLower==0 ){
1293       int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff));
1294       pLoop->nOut -= nAdjust;
1295       *pbDone = 1;
1296       WHERETRACE(0x10, ("range skip-scan regions: %u..%u  adjust=%d est=%d\n",
1297                            nLower, nUpper, nAdjust*-1, pLoop->nOut));
1298     }
1299 
1300   }else{
1301     assert( *pbDone==0 );
1302   }
1303 
1304   sqlite3ValueFree(p1);
1305   sqlite3ValueFree(p2);
1306   sqlite3ValueFree(pVal);
1307 
1308   return rc;
1309 }
1310 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1311 
1312 /*
1313 ** This function is used to estimate the number of rows that will be visited
1314 ** by scanning an index for a range of values. The range may have an upper
1315 ** bound, a lower bound, or both. The WHERE clause terms that set the upper
1316 ** and lower bounds are represented by pLower and pUpper respectively. For
1317 ** example, assuming that index p is on t1(a):
1318 **
1319 **   ... FROM t1 WHERE a > ? AND a < ? ...
1320 **                    |_____|   |_____|
1321 **                       |         |
1322 **                     pLower    pUpper
1323 **
1324 ** If either of the upper or lower bound is not present, then NULL is passed in
1325 ** place of the corresponding WhereTerm.
1326 **
1327 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index
1328 ** column subject to the range constraint. Or, equivalently, the number of
1329 ** equality constraints optimized by the proposed index scan. For example,
1330 ** assuming index p is on t1(a, b), and the SQL query is:
1331 **
1332 **   ... FROM t1 WHERE a = ? AND b > ? AND b < ? ...
1333 **
1334 ** then nEq is set to 1 (as the range restricted column, b, is the second
1335 ** left-most column of the index). Or, if the query is:
1336 **
1337 **   ... FROM t1 WHERE a > ? AND a < ? ...
1338 **
1339 ** then nEq is set to 0.
1340 **
1341 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the
1342 ** number of rows that the index scan is expected to visit without
1343 ** considering the range constraints. If nEq is 0, then *pnOut is the number of
1344 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced)
1345 ** to account for the range constraints pLower and pUpper.
1346 **
1347 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be
1348 ** used, a single range inequality reduces the search space by a factor of 4.
1349 ** and a pair of constraints (x>? AND x<?) reduces the expected number of
1350 ** rows visited by a factor of 64.
1351 */
1352 static int whereRangeScanEst(
1353   Parse *pParse,       /* Parsing & code generating context */
1354   WhereLoopBuilder *pBuilder,
1355   WhereTerm *pLower,   /* Lower bound on the range. ex: "x>123" Might be NULL */
1356   WhereTerm *pUpper,   /* Upper bound on the range. ex: "x<455" Might be NULL */
1357   WhereLoop *pLoop     /* Modify the .nOut and maybe .rRun fields */
1358 ){
1359   int rc = SQLITE_OK;
1360   int nOut = pLoop->nOut;
1361   LogEst nNew;
1362 
1363 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1364   Index *p = pLoop->u.btree.pIndex;
1365   int nEq = pLoop->u.btree.nEq;
1366 
1367   if( p->nSample>0 && nEq<p->nSampleCol ){
1368     if( nEq==pBuilder->nRecValid ){
1369       UnpackedRecord *pRec = pBuilder->pRec;
1370       tRowcnt a[2];
1371       u8 aff;
1372 
1373       /* Variable iLower will be set to the estimate of the number of rows in
1374       ** the index that are less than the lower bound of the range query. The
1375       ** lower bound being the concatenation of $P and $L, where $P is the
1376       ** key-prefix formed by the nEq values matched against the nEq left-most
1377       ** columns of the index, and $L is the value in pLower.
1378       **
1379       ** Or, if pLower is NULL or $L cannot be extracted from it (because it
1380       ** is not a simple variable or literal value), the lower bound of the
1381       ** range is $P. Due to a quirk in the way whereKeyStats() works, even
1382       ** if $L is available, whereKeyStats() is called for both ($P) and
1383       ** ($P:$L) and the larger of the two returned values is used.
1384       **
1385       ** Similarly, iUpper is to be set to the estimate of the number of rows
1386       ** less than the upper bound of the range query. Where the upper bound
1387       ** is either ($P) or ($P:$U). Again, even if $U is available, both values
1388       ** of iUpper are requested of whereKeyStats() and the smaller used.
1389       **
1390       ** The number of rows between the two bounds is then just iUpper-iLower.
1391       */
1392       tRowcnt iLower;     /* Rows less than the lower bound */
1393       tRowcnt iUpper;     /* Rows less than the upper bound */
1394       int iLwrIdx = -2;   /* aSample[] for the lower bound */
1395       int iUprIdx = -1;   /* aSample[] for the upper bound */
1396 
1397       if( pRec ){
1398         testcase( pRec->nField!=pBuilder->nRecValid );
1399         pRec->nField = pBuilder->nRecValid;
1400       }
1401       aff = sqlite3IndexColumnAffinity(pParse->db, p, nEq);
1402       assert( nEq!=p->nKeyCol || aff==SQLITE_AFF_INTEGER );
1403       /* Determine iLower and iUpper using ($P) only. */
1404       if( nEq==0 ){
1405         iLower = 0;
1406         iUpper = p->nRowEst0;
1407       }else{
1408         /* Note: this call could be optimized away - since the same values must
1409         ** have been requested when testing key $P in whereEqualScanEst().  */
1410         whereKeyStats(pParse, p, pRec, 0, a);
1411         iLower = a[0];
1412         iUpper = a[0] + a[1];
1413       }
1414 
1415       assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 );
1416       assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 );
1417       assert( p->aSortOrder!=0 );
1418       if( p->aSortOrder[nEq] ){
1419         /* The roles of pLower and pUpper are swapped for a DESC index */
1420         SWAP(WhereTerm*, pLower, pUpper);
1421       }
1422 
1423       /* If possible, improve on the iLower estimate using ($P:$L). */
1424       if( pLower ){
1425         int bOk;                    /* True if value is extracted from pExpr */
1426         Expr *pExpr = pLower->pExpr->pRight;
1427         rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq, &bOk);
1428         if( rc==SQLITE_OK && bOk ){
1429           tRowcnt iNew;
1430           iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a);
1431           iNew = a[0] + ((pLower->eOperator & (WO_GT|WO_LE)) ? a[1] : 0);
1432           if( iNew>iLower ) iLower = iNew;
1433           nOut--;
1434           pLower = 0;
1435         }
1436       }
1437 
1438       /* If possible, improve on the iUpper estimate using ($P:$U). */
1439       if( pUpper ){
1440         int bOk;                    /* True if value is extracted from pExpr */
1441         Expr *pExpr = pUpper->pExpr->pRight;
1442         rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq, &bOk);
1443         if( rc==SQLITE_OK && bOk ){
1444           tRowcnt iNew;
1445           iUprIdx = whereKeyStats(pParse, p, pRec, 1, a);
1446           iNew = a[0] + ((pUpper->eOperator & (WO_GT|WO_LE)) ? a[1] : 0);
1447           if( iNew<iUpper ) iUpper = iNew;
1448           nOut--;
1449           pUpper = 0;
1450         }
1451       }
1452 
1453       pBuilder->pRec = pRec;
1454       if( rc==SQLITE_OK ){
1455         if( iUpper>iLower ){
1456           nNew = sqlite3LogEst(iUpper - iLower);
1457           /* TUNING:  If both iUpper and iLower are derived from the same
1458           ** sample, then assume they are 4x more selective.  This brings
1459           ** the estimated selectivity more in line with what it would be
1460           ** if estimated without the use of STAT3/4 tables. */
1461           if( iLwrIdx==iUprIdx ) nNew -= 20;  assert( 20==sqlite3LogEst(4) );
1462         }else{
1463           nNew = 10;        assert( 10==sqlite3LogEst(2) );
1464         }
1465         if( nNew<nOut ){
1466           nOut = nNew;
1467         }
1468         WHERETRACE(0x10, ("STAT4 range scan: %u..%u  est=%d\n",
1469                            (u32)iLower, (u32)iUpper, nOut));
1470       }
1471     }else{
1472       int bDone = 0;
1473       rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone);
1474       if( bDone ) return rc;
1475     }
1476   }
1477 #else
1478   UNUSED_PARAMETER(pParse);
1479   UNUSED_PARAMETER(pBuilder);
1480   assert( pLower || pUpper );
1481 #endif
1482   assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 );
1483   nNew = whereRangeAdjust(pLower, nOut);
1484   nNew = whereRangeAdjust(pUpper, nNew);
1485 
1486   /* TUNING: If there is both an upper and lower limit and neither limit
1487   ** has an application-defined likelihood(), assume the range is
1488   ** reduced by an additional 75%. This means that, by default, an open-ended
1489   ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the
1490   ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to
1491   ** match 1/64 of the index. */
1492   if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){
1493     nNew -= 20;
1494   }
1495 
1496   nOut -= (pLower!=0) + (pUpper!=0);
1497   if( nNew<10 ) nNew = 10;
1498   if( nNew<nOut ) nOut = nNew;
1499 #if defined(WHERETRACE_ENABLED)
1500   if( pLoop->nOut>nOut ){
1501     WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n",
1502                     pLoop->nOut, nOut));
1503   }
1504 #endif
1505   pLoop->nOut = (LogEst)nOut;
1506   return rc;
1507 }
1508 
1509 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1510 /*
1511 ** Estimate the number of rows that will be returned based on
1512 ** an equality constraint x=VALUE and where that VALUE occurs in
1513 ** the histogram data.  This only works when x is the left-most
1514 ** column of an index and sqlite_stat3 histogram data is available
1515 ** for that index.  When pExpr==NULL that means the constraint is
1516 ** "x IS NULL" instead of "x=VALUE".
1517 **
1518 ** Write the estimated row count into *pnRow and return SQLITE_OK.
1519 ** If unable to make an estimate, leave *pnRow unchanged and return
1520 ** non-zero.
1521 **
1522 ** This routine can fail if it is unable to load a collating sequence
1523 ** required for string comparison, or if unable to allocate memory
1524 ** for a UTF conversion required for comparison.  The error is stored
1525 ** in the pParse structure.
1526 */
1527 static int whereEqualScanEst(
1528   Parse *pParse,       /* Parsing & code generating context */
1529   WhereLoopBuilder *pBuilder,
1530   Expr *pExpr,         /* Expression for VALUE in the x=VALUE constraint */
1531   tRowcnt *pnRow       /* Write the revised row estimate here */
1532 ){
1533   Index *p = pBuilder->pNew->u.btree.pIndex;
1534   int nEq = pBuilder->pNew->u.btree.nEq;
1535   UnpackedRecord *pRec = pBuilder->pRec;
1536   u8 aff;                   /* Column affinity */
1537   int rc;                   /* Subfunction return code */
1538   tRowcnt a[2];             /* Statistics */
1539   int bOk;
1540 
1541   assert( nEq>=1 );
1542   assert( nEq<=p->nColumn );
1543   assert( p->aSample!=0 );
1544   assert( p->nSample>0 );
1545   assert( pBuilder->nRecValid<nEq );
1546 
1547   /* If values are not available for all fields of the index to the left
1548   ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */
1549   if( pBuilder->nRecValid<(nEq-1) ){
1550     return SQLITE_NOTFOUND;
1551   }
1552 
1553   /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue()
1554   ** below would return the same value.  */
1555   if( nEq>=p->nColumn ){
1556     *pnRow = 1;
1557     return SQLITE_OK;
1558   }
1559 
1560   aff = sqlite3IndexColumnAffinity(pParse->db, p, nEq-1);
1561   rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, aff, nEq-1, &bOk);
1562   pBuilder->pRec = pRec;
1563   if( rc!=SQLITE_OK ) return rc;
1564   if( bOk==0 ) return SQLITE_NOTFOUND;
1565   pBuilder->nRecValid = nEq;
1566 
1567   whereKeyStats(pParse, p, pRec, 0, a);
1568   WHERETRACE(0x10,("equality scan regions %s(%d): %d\n",
1569                    p->zName, nEq-1, (int)a[1]));
1570   *pnRow = a[1];
1571 
1572   return rc;
1573 }
1574 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1575 
1576 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
1577 /*
1578 ** Estimate the number of rows that will be returned based on
1579 ** an IN constraint where the right-hand side of the IN operator
1580 ** is a list of values.  Example:
1581 **
1582 **        WHERE x IN (1,2,3,4)
1583 **
1584 ** Write the estimated row count into *pnRow and return SQLITE_OK.
1585 ** If unable to make an estimate, leave *pnRow unchanged and return
1586 ** non-zero.
1587 **
1588 ** This routine can fail if it is unable to load a collating sequence
1589 ** required for string comparison, or if unable to allocate memory
1590 ** for a UTF conversion required for comparison.  The error is stored
1591 ** in the pParse structure.
1592 */
1593 static int whereInScanEst(
1594   Parse *pParse,       /* Parsing & code generating context */
1595   WhereLoopBuilder *pBuilder,
1596   ExprList *pList,     /* The value list on the RHS of "x IN (v1,v2,v3,...)" */
1597   tRowcnt *pnRow       /* Write the revised row estimate here */
1598 ){
1599   Index *p = pBuilder->pNew->u.btree.pIndex;
1600   i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]);
1601   int nRecValid = pBuilder->nRecValid;
1602   int rc = SQLITE_OK;     /* Subfunction return code */
1603   tRowcnt nEst;           /* Number of rows for a single term */
1604   tRowcnt nRowEst = 0;    /* New estimate of the number of rows */
1605   int i;                  /* Loop counter */
1606 
1607   assert( p->aSample!=0 );
1608   for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){
1609     nEst = nRow0;
1610     rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst);
1611     nRowEst += nEst;
1612     pBuilder->nRecValid = nRecValid;
1613   }
1614 
1615   if( rc==SQLITE_OK ){
1616     if( nRowEst > nRow0 ) nRowEst = nRow0;
1617     *pnRow = nRowEst;
1618     WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst));
1619   }
1620   assert( pBuilder->nRecValid==nRecValid );
1621   return rc;
1622 }
1623 #endif /* SQLITE_ENABLE_STAT3_OR_STAT4 */
1624 
1625 
1626 #ifdef WHERETRACE_ENABLED
1627 /*
1628 ** Print the content of a WhereTerm object
1629 */
1630 static void whereTermPrint(WhereTerm *pTerm, int iTerm){
1631   if( pTerm==0 ){
1632     sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm);
1633   }else{
1634     char zType[4];
1635     memcpy(zType, "...", 4);
1636     if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V';
1637     if( pTerm->eOperator & WO_EQUIV  ) zType[1] = 'E';
1638     if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L';
1639     sqlite3DebugPrintf(
1640        "TERM-%-3d %p %s cursor=%-3d prob=%-3d op=0x%03x wtFlags=0x%04x\n",
1641        iTerm, pTerm, zType, pTerm->leftCursor, pTerm->truthProb,
1642        pTerm->eOperator, pTerm->wtFlags);
1643     sqlite3TreeViewExpr(0, pTerm->pExpr, 0);
1644   }
1645 }
1646 #endif
1647 
1648 #ifdef WHERETRACE_ENABLED
1649 /*
1650 ** Print a WhereLoop object for debugging purposes
1651 */
1652 static void whereLoopPrint(WhereLoop *p, WhereClause *pWC){
1653   WhereInfo *pWInfo = pWC->pWInfo;
1654   int nb = 1+(pWInfo->pTabList->nSrc+7)/8;
1655   struct SrcList_item *pItem = pWInfo->pTabList->a + p->iTab;
1656   Table *pTab = pItem->pTab;
1657   sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId,
1658                      p->iTab, nb, p->maskSelf, nb, p->prereq);
1659   sqlite3DebugPrintf(" %12s",
1660                      pItem->zAlias ? pItem->zAlias : pTab->zName);
1661   if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){
1662     const char *zName;
1663     if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){
1664       if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){
1665         int i = sqlite3Strlen30(zName) - 1;
1666         while( zName[i]!='_' ) i--;
1667         zName += i;
1668       }
1669       sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq);
1670     }else{
1671       sqlite3DebugPrintf("%20s","");
1672     }
1673   }else{
1674     char *z;
1675     if( p->u.vtab.idxStr ){
1676       z = sqlite3_mprintf("(%d,\"%s\",%x)",
1677                 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask);
1678     }else{
1679       z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask);
1680     }
1681     sqlite3DebugPrintf(" %-19s", z);
1682     sqlite3_free(z);
1683   }
1684   if( p->wsFlags & WHERE_SKIPSCAN ){
1685     sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip);
1686   }else{
1687     sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm);
1688   }
1689   sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut);
1690   if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){
1691     int i;
1692     for(i=0; i<p->nLTerm; i++){
1693       whereTermPrint(p->aLTerm[i], i);
1694     }
1695   }
1696 }
1697 #endif
1698 
1699 /*
1700 ** Convert bulk memory into a valid WhereLoop that can be passed
1701 ** to whereLoopClear harmlessly.
1702 */
1703 static void whereLoopInit(WhereLoop *p){
1704   p->aLTerm = p->aLTermSpace;
1705   p->nLTerm = 0;
1706   p->nLSlot = ArraySize(p->aLTermSpace);
1707   p->wsFlags = 0;
1708 }
1709 
1710 /*
1711 ** Clear the WhereLoop.u union.  Leave WhereLoop.pLTerm intact.
1712 */
1713 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){
1714   if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){
1715     if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){
1716       sqlite3_free(p->u.vtab.idxStr);
1717       p->u.vtab.needFree = 0;
1718       p->u.vtab.idxStr = 0;
1719     }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){
1720       sqlite3DbFree(db, p->u.btree.pIndex->zColAff);
1721       sqlite3DbFree(db, p->u.btree.pIndex);
1722       p->u.btree.pIndex = 0;
1723     }
1724   }
1725 }
1726 
1727 /*
1728 ** Deallocate internal memory used by a WhereLoop object
1729 */
1730 static void whereLoopClear(sqlite3 *db, WhereLoop *p){
1731   if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFree(db, p->aLTerm);
1732   whereLoopClearUnion(db, p);
1733   whereLoopInit(p);
1734 }
1735 
1736 /*
1737 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n.
1738 */
1739 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){
1740   WhereTerm **paNew;
1741   if( p->nLSlot>=n ) return SQLITE_OK;
1742   n = (n+7)&~7;
1743   paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n);
1744   if( paNew==0 ) return SQLITE_NOMEM_BKPT;
1745   memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot);
1746   if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFree(db, p->aLTerm);
1747   p->aLTerm = paNew;
1748   p->nLSlot = n;
1749   return SQLITE_OK;
1750 }
1751 
1752 /*
1753 ** Transfer content from the second pLoop into the first.
1754 */
1755 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){
1756   whereLoopClearUnion(db, pTo);
1757   if( whereLoopResize(db, pTo, pFrom->nLTerm) ){
1758     memset(&pTo->u, 0, sizeof(pTo->u));
1759     return SQLITE_NOMEM_BKPT;
1760   }
1761   memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ);
1762   memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0]));
1763   if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){
1764     pFrom->u.vtab.needFree = 0;
1765   }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){
1766     pFrom->u.btree.pIndex = 0;
1767   }
1768   return SQLITE_OK;
1769 }
1770 
1771 /*
1772 ** Delete a WhereLoop object
1773 */
1774 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){
1775   whereLoopClear(db, p);
1776   sqlite3DbFree(db, p);
1777 }
1778 
1779 /*
1780 ** Free a WhereInfo structure
1781 */
1782 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){
1783   if( ALWAYS(pWInfo) ){
1784     int i;
1785     for(i=0; i<pWInfo->nLevel; i++){
1786       WhereLevel *pLevel = &pWInfo->a[i];
1787       if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){
1788         sqlite3DbFree(db, pLevel->u.in.aInLoop);
1789       }
1790     }
1791     sqlite3WhereClauseClear(&pWInfo->sWC);
1792     while( pWInfo->pLoops ){
1793       WhereLoop *p = pWInfo->pLoops;
1794       pWInfo->pLoops = p->pNextLoop;
1795       whereLoopDelete(db, p);
1796     }
1797     sqlite3DbFree(db, pWInfo);
1798   }
1799 }
1800 
1801 /*
1802 ** Return TRUE if all of the following are true:
1803 **
1804 **   (1)  X has the same or lower cost that Y
1805 **   (2)  X is a proper subset of Y
1806 **   (3)  X skips at least as many columns as Y
1807 **
1808 ** By "proper subset" we mean that X uses fewer WHERE clause terms
1809 ** than Y and that every WHERE clause term used by X is also used
1810 ** by Y.
1811 **
1812 ** If X is a proper subset of Y then Y is a better choice and ought
1813 ** to have a lower cost.  This routine returns TRUE when that cost
1814 ** relationship is inverted and needs to be adjusted.  The third rule
1815 ** was added because if X uses skip-scan less than Y it still might
1816 ** deserve a lower cost even if it is a proper subset of Y.
1817 */
1818 static int whereLoopCheaperProperSubset(
1819   const WhereLoop *pX,       /* First WhereLoop to compare */
1820   const WhereLoop *pY        /* Compare against this WhereLoop */
1821 ){
1822   int i, j;
1823   if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){
1824     return 0; /* X is not a subset of Y */
1825   }
1826   if( pY->nSkip > pX->nSkip ) return 0;
1827   if( pX->rRun >= pY->rRun ){
1828     if( pX->rRun > pY->rRun ) return 0;    /* X costs more than Y */
1829     if( pX->nOut > pY->nOut ) return 0;    /* X costs more than Y */
1830   }
1831   for(i=pX->nLTerm-1; i>=0; i--){
1832     if( pX->aLTerm[i]==0 ) continue;
1833     for(j=pY->nLTerm-1; j>=0; j--){
1834       if( pY->aLTerm[j]==pX->aLTerm[i] ) break;
1835     }
1836     if( j<0 ) return 0;  /* X not a subset of Y since term X[i] not used by Y */
1837   }
1838   return 1;  /* All conditions meet */
1839 }
1840 
1841 /*
1842 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so
1843 ** that:
1844 **
1845 **   (1) pTemplate costs less than any other WhereLoops that are a proper
1846 **       subset of pTemplate
1847 **
1848 **   (2) pTemplate costs more than any other WhereLoops for which pTemplate
1849 **       is a proper subset.
1850 **
1851 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer
1852 ** WHERE clause terms than Y and that every WHERE clause term used by X is
1853 ** also used by Y.
1854 */
1855 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){
1856   if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return;
1857   for(; p; p=p->pNextLoop){
1858     if( p->iTab!=pTemplate->iTab ) continue;
1859     if( (p->wsFlags & WHERE_INDEXED)==0 ) continue;
1860     if( whereLoopCheaperProperSubset(p, pTemplate) ){
1861       /* Adjust pTemplate cost downward so that it is cheaper than its
1862       ** subset p. */
1863       WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
1864                        pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1));
1865       pTemplate->rRun = p->rRun;
1866       pTemplate->nOut = p->nOut - 1;
1867     }else if( whereLoopCheaperProperSubset(pTemplate, p) ){
1868       /* Adjust pTemplate cost upward so that it is costlier than p since
1869       ** pTemplate is a proper subset of p */
1870       WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n",
1871                        pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1));
1872       pTemplate->rRun = p->rRun;
1873       pTemplate->nOut = p->nOut + 1;
1874     }
1875   }
1876 }
1877 
1878 /*
1879 ** Search the list of WhereLoops in *ppPrev looking for one that can be
1880 ** supplanted by pTemplate.
1881 **
1882 ** Return NULL if the WhereLoop list contains an entry that can supplant
1883 ** pTemplate, in other words if pTemplate does not belong on the list.
1884 **
1885 ** If pX is a WhereLoop that pTemplate can supplant, then return the
1886 ** link that points to pX.
1887 **
1888 ** If pTemplate cannot supplant any existing element of the list but needs
1889 ** to be added to the list, then return a pointer to the tail of the list.
1890 */
1891 static WhereLoop **whereLoopFindLesser(
1892   WhereLoop **ppPrev,
1893   const WhereLoop *pTemplate
1894 ){
1895   WhereLoop *p;
1896   for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){
1897     if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){
1898       /* If either the iTab or iSortIdx values for two WhereLoop are different
1899       ** then those WhereLoops need to be considered separately.  Neither is
1900       ** a candidate to replace the other. */
1901       continue;
1902     }
1903     /* In the current implementation, the rSetup value is either zero
1904     ** or the cost of building an automatic index (NlogN) and the NlogN
1905     ** is the same for compatible WhereLoops. */
1906     assert( p->rSetup==0 || pTemplate->rSetup==0
1907                  || p->rSetup==pTemplate->rSetup );
1908 
1909     /* whereLoopAddBtree() always generates and inserts the automatic index
1910     ** case first.  Hence compatible candidate WhereLoops never have a larger
1911     ** rSetup. Call this SETUP-INVARIANT */
1912     assert( p->rSetup>=pTemplate->rSetup );
1913 
1914     /* Any loop using an appliation-defined index (or PRIMARY KEY or
1915     ** UNIQUE constraint) with one or more == constraints is better
1916     ** than an automatic index. Unless it is a skip-scan. */
1917     if( (p->wsFlags & WHERE_AUTO_INDEX)!=0
1918      && (pTemplate->nSkip)==0
1919      && (pTemplate->wsFlags & WHERE_INDEXED)!=0
1920      && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0
1921      && (p->prereq & pTemplate->prereq)==pTemplate->prereq
1922     ){
1923       break;
1924     }
1925 
1926     /* If existing WhereLoop p is better than pTemplate, pTemplate can be
1927     ** discarded.  WhereLoop p is better if:
1928     **   (1)  p has no more dependencies than pTemplate, and
1929     **   (2)  p has an equal or lower cost than pTemplate
1930     */
1931     if( (p->prereq & pTemplate->prereq)==p->prereq    /* (1)  */
1932      && p->rSetup<=pTemplate->rSetup                  /* (2a) */
1933      && p->rRun<=pTemplate->rRun                      /* (2b) */
1934      && p->nOut<=pTemplate->nOut                      /* (2c) */
1935     ){
1936       return 0;  /* Discard pTemplate */
1937     }
1938 
1939     /* If pTemplate is always better than p, then cause p to be overwritten
1940     ** with pTemplate.  pTemplate is better than p if:
1941     **   (1)  pTemplate has no more dependences than p, and
1942     **   (2)  pTemplate has an equal or lower cost than p.
1943     */
1944     if( (p->prereq & pTemplate->prereq)==pTemplate->prereq   /* (1)  */
1945      && p->rRun>=pTemplate->rRun                             /* (2a) */
1946      && p->nOut>=pTemplate->nOut                             /* (2b) */
1947     ){
1948       assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */
1949       break;   /* Cause p to be overwritten by pTemplate */
1950     }
1951   }
1952   return ppPrev;
1953 }
1954 
1955 /*
1956 ** Insert or replace a WhereLoop entry using the template supplied.
1957 **
1958 ** An existing WhereLoop entry might be overwritten if the new template
1959 ** is better and has fewer dependencies.  Or the template will be ignored
1960 ** and no insert will occur if an existing WhereLoop is faster and has
1961 ** fewer dependencies than the template.  Otherwise a new WhereLoop is
1962 ** added based on the template.
1963 **
1964 ** If pBuilder->pOrSet is not NULL then we care about only the
1965 ** prerequisites and rRun and nOut costs of the N best loops.  That
1966 ** information is gathered in the pBuilder->pOrSet object.  This special
1967 ** processing mode is used only for OR clause processing.
1968 **
1969 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we
1970 ** still might overwrite similar loops with the new template if the
1971 ** new template is better.  Loops may be overwritten if the following
1972 ** conditions are met:
1973 **
1974 **    (1)  They have the same iTab.
1975 **    (2)  They have the same iSortIdx.
1976 **    (3)  The template has same or fewer dependencies than the current loop
1977 **    (4)  The template has the same or lower cost than the current loop
1978 */
1979 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){
1980   WhereLoop **ppPrev, *p;
1981   WhereInfo *pWInfo = pBuilder->pWInfo;
1982   sqlite3 *db = pWInfo->pParse->db;
1983   int rc;
1984 
1985   /* If pBuilder->pOrSet is defined, then only keep track of the costs
1986   ** and prereqs.
1987   */
1988   if( pBuilder->pOrSet!=0 ){
1989     if( pTemplate->nLTerm ){
1990 #if WHERETRACE_ENABLED
1991       u16 n = pBuilder->pOrSet->n;
1992       int x =
1993 #endif
1994       whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun,
1995                                     pTemplate->nOut);
1996 #if WHERETRACE_ENABLED /* 0x8 */
1997       if( sqlite3WhereTrace & 0x8 ){
1998         sqlite3DebugPrintf(x?"   or-%d:  ":"   or-X:  ", n);
1999         whereLoopPrint(pTemplate, pBuilder->pWC);
2000       }
2001 #endif
2002     }
2003     return SQLITE_OK;
2004   }
2005 
2006   /* Look for an existing WhereLoop to replace with pTemplate
2007   */
2008   whereLoopAdjustCost(pWInfo->pLoops, pTemplate);
2009   ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate);
2010 
2011   if( ppPrev==0 ){
2012     /* There already exists a WhereLoop on the list that is better
2013     ** than pTemplate, so just ignore pTemplate */
2014 #if WHERETRACE_ENABLED /* 0x8 */
2015     if( sqlite3WhereTrace & 0x8 ){
2016       sqlite3DebugPrintf("   skip: ");
2017       whereLoopPrint(pTemplate, pBuilder->pWC);
2018     }
2019 #endif
2020     return SQLITE_OK;
2021   }else{
2022     p = *ppPrev;
2023   }
2024 
2025   /* If we reach this point it means that either p[] should be overwritten
2026   ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new
2027   ** WhereLoop and insert it.
2028   */
2029 #if WHERETRACE_ENABLED /* 0x8 */
2030   if( sqlite3WhereTrace & 0x8 ){
2031     if( p!=0 ){
2032       sqlite3DebugPrintf("replace: ");
2033       whereLoopPrint(p, pBuilder->pWC);
2034     }
2035     sqlite3DebugPrintf("    add: ");
2036     whereLoopPrint(pTemplate, pBuilder->pWC);
2037   }
2038 #endif
2039   if( p==0 ){
2040     /* Allocate a new WhereLoop to add to the end of the list */
2041     *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop));
2042     if( p==0 ) return SQLITE_NOMEM_BKPT;
2043     whereLoopInit(p);
2044     p->pNextLoop = 0;
2045   }else{
2046     /* We will be overwriting WhereLoop p[].  But before we do, first
2047     ** go through the rest of the list and delete any other entries besides
2048     ** p[] that are also supplated by pTemplate */
2049     WhereLoop **ppTail = &p->pNextLoop;
2050     WhereLoop *pToDel;
2051     while( *ppTail ){
2052       ppTail = whereLoopFindLesser(ppTail, pTemplate);
2053       if( ppTail==0 ) break;
2054       pToDel = *ppTail;
2055       if( pToDel==0 ) break;
2056       *ppTail = pToDel->pNextLoop;
2057 #if WHERETRACE_ENABLED /* 0x8 */
2058       if( sqlite3WhereTrace & 0x8 ){
2059         sqlite3DebugPrintf(" delete: ");
2060         whereLoopPrint(pToDel, pBuilder->pWC);
2061       }
2062 #endif
2063       whereLoopDelete(db, pToDel);
2064     }
2065   }
2066   rc = whereLoopXfer(db, p, pTemplate);
2067   if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){
2068     Index *pIndex = p->u.btree.pIndex;
2069     if( pIndex && pIndex->tnum==0 ){
2070       p->u.btree.pIndex = 0;
2071     }
2072   }
2073   return rc;
2074 }
2075 
2076 /*
2077 ** Adjust the WhereLoop.nOut value downward to account for terms of the
2078 ** WHERE clause that reference the loop but which are not used by an
2079 ** index.
2080 *
2081 ** For every WHERE clause term that is not used by the index
2082 ** and which has a truth probability assigned by one of the likelihood(),
2083 ** likely(), or unlikely() SQL functions, reduce the estimated number
2084 ** of output rows by the probability specified.
2085 **
2086 ** TUNING:  For every WHERE clause term that is not used by the index
2087 ** and which does not have an assigned truth probability, heuristics
2088 ** described below are used to try to estimate the truth probability.
2089 ** TODO --> Perhaps this is something that could be improved by better
2090 ** table statistics.
2091 **
2092 ** Heuristic 1:  Estimate the truth probability as 93.75%.  The 93.75%
2093 ** value corresponds to -1 in LogEst notation, so this means decrement
2094 ** the WhereLoop.nOut field for every such WHERE clause term.
2095 **
2096 ** Heuristic 2:  If there exists one or more WHERE clause terms of the
2097 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the
2098 ** final output row estimate is no greater than 1/4 of the total number
2099 ** of rows in the table.  In other words, assume that x==EXPR will filter
2100 ** out at least 3 out of 4 rows.  If EXPR is -1 or 0 or 1, then maybe the
2101 ** "x" column is boolean or else -1 or 0 or 1 is a common default value
2102 ** on the "x" column and so in that case only cap the output row estimate
2103 ** at 1/2 instead of 1/4.
2104 */
2105 static void whereLoopOutputAdjust(
2106   WhereClause *pWC,      /* The WHERE clause */
2107   WhereLoop *pLoop,      /* The loop to adjust downward */
2108   LogEst nRow            /* Number of rows in the entire table */
2109 ){
2110   WhereTerm *pTerm, *pX;
2111   Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf);
2112   int i, j, k;
2113   LogEst iReduce = 0;    /* pLoop->nOut should not exceed nRow-iReduce */
2114 
2115   assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 );
2116   for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){
2117     if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break;
2118     if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue;
2119     if( (pTerm->prereqAll & notAllowed)!=0 ) continue;
2120     for(j=pLoop->nLTerm-1; j>=0; j--){
2121       pX = pLoop->aLTerm[j];
2122       if( pX==0 ) continue;
2123       if( pX==pTerm ) break;
2124       if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break;
2125     }
2126     if( j<0 ){
2127       if( pTerm->truthProb<=0 ){
2128         /* If a truth probability is specified using the likelihood() hints,
2129         ** then use the probability provided by the application. */
2130         pLoop->nOut += pTerm->truthProb;
2131       }else{
2132         /* In the absence of explicit truth probabilities, use heuristics to
2133         ** guess a reasonable truth probability. */
2134         pLoop->nOut--;
2135         if( pTerm->eOperator&(WO_EQ|WO_IS) ){
2136           Expr *pRight = pTerm->pExpr->pRight;
2137           testcase( pTerm->pExpr->op==TK_IS );
2138           if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){
2139             k = 10;
2140           }else{
2141             k = 20;
2142           }
2143           if( iReduce<k ) iReduce = k;
2144         }
2145       }
2146     }
2147   }
2148   if( pLoop->nOut > nRow-iReduce )  pLoop->nOut = nRow - iReduce;
2149 }
2150 
2151 /*
2152 ** Adjust the cost C by the costMult facter T.  This only occurs if
2153 ** compiled with -DSQLITE_ENABLE_COSTMULT
2154 */
2155 #ifdef SQLITE_ENABLE_COSTMULT
2156 # define ApplyCostMultiplier(C,T)  C += T
2157 #else
2158 # define ApplyCostMultiplier(C,T)
2159 #endif
2160 
2161 /*
2162 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the
2163 ** index pIndex. Try to match one more.
2164 **
2165 ** When this function is called, pBuilder->pNew->nOut contains the
2166 ** number of rows expected to be visited by filtering using the nEq
2167 ** terms only. If it is modified, this value is restored before this
2168 ** function returns.
2169 **
2170 ** If pProbe->tnum==0, that means pIndex is a fake index used for the
2171 ** INTEGER PRIMARY KEY.
2172 */
2173 static int whereLoopAddBtreeIndex(
2174   WhereLoopBuilder *pBuilder,     /* The WhereLoop factory */
2175   struct SrcList_item *pSrc,      /* FROM clause term being analyzed */
2176   Index *pProbe,                  /* An index on pSrc */
2177   LogEst nInMul                   /* log(Number of iterations due to IN) */
2178 ){
2179   WhereInfo *pWInfo = pBuilder->pWInfo;  /* WHERE analyse context */
2180   Parse *pParse = pWInfo->pParse;        /* Parsing context */
2181   sqlite3 *db = pParse->db;       /* Database connection malloc context */
2182   WhereLoop *pNew;                /* Template WhereLoop under construction */
2183   WhereTerm *pTerm;               /* A WhereTerm under consideration */
2184   int opMask;                     /* Valid operators for constraints */
2185   WhereScan scan;                 /* Iterator for WHERE terms */
2186   Bitmask saved_prereq;           /* Original value of pNew->prereq */
2187   u16 saved_nLTerm;               /* Original value of pNew->nLTerm */
2188   u16 saved_nEq;                  /* Original value of pNew->u.btree.nEq */
2189   u16 saved_nSkip;                /* Original value of pNew->nSkip */
2190   u32 saved_wsFlags;              /* Original value of pNew->wsFlags */
2191   LogEst saved_nOut;              /* Original value of pNew->nOut */
2192   int rc = SQLITE_OK;             /* Return code */
2193   LogEst rSize;                   /* Number of rows in the table */
2194   LogEst rLogSize;                /* Logarithm of table size */
2195   WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */
2196 
2197   pNew = pBuilder->pNew;
2198   if( db->mallocFailed ) return SQLITE_NOMEM_BKPT;
2199 
2200   assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 );
2201   assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 );
2202   if( pNew->wsFlags & WHERE_BTM_LIMIT ){
2203     opMask = WO_LT|WO_LE;
2204   }else if( /*pProbe->tnum<=0 ||*/ (pSrc->fg.jointype & JT_LEFT)!=0 ){
2205     opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE;
2206   }else{
2207     opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS;
2208   }
2209   if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE);
2210 
2211   assert( pNew->u.btree.nEq<pProbe->nColumn );
2212 
2213   saved_nEq = pNew->u.btree.nEq;
2214   saved_nSkip = pNew->nSkip;
2215   saved_nLTerm = pNew->nLTerm;
2216   saved_wsFlags = pNew->wsFlags;
2217   saved_prereq = pNew->prereq;
2218   saved_nOut = pNew->nOut;
2219   pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq,
2220                         opMask, pProbe);
2221   pNew->rSetup = 0;
2222   rSize = pProbe->aiRowLogEst[0];
2223   rLogSize = estLog(rSize);
2224   for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){
2225     u16 eOp = pTerm->eOperator;   /* Shorthand for pTerm->eOperator */
2226     LogEst rCostIdx;
2227     LogEst nOutUnadjusted;        /* nOut before IN() and WHERE adjustments */
2228     int nIn = 0;
2229 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2230     int nRecValid = pBuilder->nRecValid;
2231 #endif
2232     if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0)
2233      && indexColumnNotNull(pProbe, saved_nEq)
2234     ){
2235       continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */
2236     }
2237     if( pTerm->prereqRight & pNew->maskSelf ) continue;
2238 
2239     /* Do not allow the upper bound of a LIKE optimization range constraint
2240     ** to mix with a lower range bound from some other source */
2241     if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue;
2242 
2243     pNew->wsFlags = saved_wsFlags;
2244     pNew->u.btree.nEq = saved_nEq;
2245     pNew->nLTerm = saved_nLTerm;
2246     if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */
2247     pNew->aLTerm[pNew->nLTerm++] = pTerm;
2248     pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf;
2249 
2250     assert( nInMul==0
2251         || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0
2252         || (pNew->wsFlags & WHERE_COLUMN_IN)!=0
2253         || (pNew->wsFlags & WHERE_SKIPSCAN)!=0
2254     );
2255 
2256     if( eOp & WO_IN ){
2257       Expr *pExpr = pTerm->pExpr;
2258       pNew->wsFlags |= WHERE_COLUMN_IN;
2259       if( ExprHasProperty(pExpr, EP_xIsSelect) ){
2260         /* "x IN (SELECT ...)":  TUNING: the SELECT returns 25 rows */
2261         nIn = 46;  assert( 46==sqlite3LogEst(25) );
2262       }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){
2263         /* "x IN (value, value, ...)" */
2264         nIn = sqlite3LogEst(pExpr->x.pList->nExpr);
2265       }
2266       assert( nIn>0 );  /* RHS always has 2 or more terms...  The parser
2267                         ** changes "x IN (?)" into "x=?". */
2268 
2269     }else if( eOp & (WO_EQ|WO_IS) ){
2270       int iCol = pProbe->aiColumn[saved_nEq];
2271       pNew->wsFlags |= WHERE_COLUMN_EQ;
2272       assert( saved_nEq==pNew->u.btree.nEq );
2273       if( iCol==XN_ROWID
2274        || (iCol>0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1)
2275       ){
2276         if( iCol>=0 && pProbe->uniqNotNull==0 ){
2277           pNew->wsFlags |= WHERE_UNQ_WANTED;
2278         }else{
2279           pNew->wsFlags |= WHERE_ONEROW;
2280         }
2281       }
2282     }else if( eOp & WO_ISNULL ){
2283       pNew->wsFlags |= WHERE_COLUMN_NULL;
2284     }else if( eOp & (WO_GT|WO_GE) ){
2285       testcase( eOp & WO_GT );
2286       testcase( eOp & WO_GE );
2287       pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT;
2288       pBtm = pTerm;
2289       pTop = 0;
2290       if( pTerm->wtFlags & TERM_LIKEOPT ){
2291         /* Range contraints that come from the LIKE optimization are
2292         ** always used in pairs. */
2293         pTop = &pTerm[1];
2294         assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm );
2295         assert( pTop->wtFlags & TERM_LIKEOPT );
2296         assert( pTop->eOperator==WO_LT );
2297         if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */
2298         pNew->aLTerm[pNew->nLTerm++] = pTop;
2299         pNew->wsFlags |= WHERE_TOP_LIMIT;
2300       }
2301     }else{
2302       assert( eOp & (WO_LT|WO_LE) );
2303       testcase( eOp & WO_LT );
2304       testcase( eOp & WO_LE );
2305       pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT;
2306       pTop = pTerm;
2307       pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ?
2308                      pNew->aLTerm[pNew->nLTerm-2] : 0;
2309     }
2310 
2311     /* At this point pNew->nOut is set to the number of rows expected to
2312     ** be visited by the index scan before considering term pTerm, or the
2313     ** values of nIn and nInMul. In other words, assuming that all
2314     ** "x IN(...)" terms are replaced with "x = ?". This block updates
2315     ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul).  */
2316     assert( pNew->nOut==saved_nOut );
2317     if( pNew->wsFlags & WHERE_COLUMN_RANGE ){
2318       /* Adjust nOut using stat3/stat4 data. Or, if there is no stat3/stat4
2319       ** data, using some other estimate.  */
2320       whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew);
2321     }else{
2322       int nEq = ++pNew->u.btree.nEq;
2323       assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) );
2324 
2325       assert( pNew->nOut==saved_nOut );
2326       if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){
2327         assert( (eOp & WO_IN) || nIn==0 );
2328         testcase( eOp & WO_IN );
2329         pNew->nOut += pTerm->truthProb;
2330         pNew->nOut -= nIn;
2331       }else{
2332 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2333         tRowcnt nOut = 0;
2334         if( nInMul==0
2335          && pProbe->nSample
2336          && pNew->u.btree.nEq<=pProbe->nSampleCol
2337          && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect))
2338         ){
2339           Expr *pExpr = pTerm->pExpr;
2340           if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){
2341             testcase( eOp & WO_EQ );
2342             testcase( eOp & WO_IS );
2343             testcase( eOp & WO_ISNULL );
2344             rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut);
2345           }else{
2346             rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut);
2347           }
2348           if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK;
2349           if( rc!=SQLITE_OK ) break;          /* Jump out of the pTerm loop */
2350           if( nOut ){
2351             pNew->nOut = sqlite3LogEst(nOut);
2352             if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut;
2353             pNew->nOut -= nIn;
2354           }
2355         }
2356         if( nOut==0 )
2357 #endif
2358         {
2359           pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]);
2360           if( eOp & WO_ISNULL ){
2361             /* TUNING: If there is no likelihood() value, assume that a
2362             ** "col IS NULL" expression matches twice as many rows
2363             ** as (col=?). */
2364             pNew->nOut += 10;
2365           }
2366         }
2367       }
2368     }
2369 
2370     /* Set rCostIdx to the cost of visiting selected rows in index. Add
2371     ** it to pNew->rRun, which is currently set to the cost of the index
2372     ** seek only. Then, if this is a non-covering index, add the cost of
2373     ** visiting the rows in the main table.  */
2374     rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow;
2375     pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx);
2376     if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){
2377       pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16);
2378     }
2379     ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult);
2380 
2381     nOutUnadjusted = pNew->nOut;
2382     pNew->rRun += nInMul + nIn;
2383     pNew->nOut += nInMul + nIn;
2384     whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize);
2385     rc = whereLoopInsert(pBuilder, pNew);
2386 
2387     if( pNew->wsFlags & WHERE_COLUMN_RANGE ){
2388       pNew->nOut = saved_nOut;
2389     }else{
2390       pNew->nOut = nOutUnadjusted;
2391     }
2392 
2393     if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0
2394      && pNew->u.btree.nEq<pProbe->nColumn
2395     ){
2396       whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn);
2397     }
2398     pNew->nOut = saved_nOut;
2399 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2400     pBuilder->nRecValid = nRecValid;
2401 #endif
2402   }
2403   pNew->prereq = saved_prereq;
2404   pNew->u.btree.nEq = saved_nEq;
2405   pNew->nSkip = saved_nSkip;
2406   pNew->wsFlags = saved_wsFlags;
2407   pNew->nOut = saved_nOut;
2408   pNew->nLTerm = saved_nLTerm;
2409 
2410   /* Consider using a skip-scan if there are no WHERE clause constraints
2411   ** available for the left-most terms of the index, and if the average
2412   ** number of repeats in the left-most terms is at least 18.
2413   **
2414   ** The magic number 18 is selected on the basis that scanning 17 rows
2415   ** is almost always quicker than an index seek (even though if the index
2416   ** contains fewer than 2^17 rows we assume otherwise in other parts of
2417   ** the code). And, even if it is not, it should not be too much slower.
2418   ** On the other hand, the extra seeks could end up being significantly
2419   ** more expensive.  */
2420   assert( 42==sqlite3LogEst(18) );
2421   if( saved_nEq==saved_nSkip
2422    && saved_nEq+1<pProbe->nKeyCol
2423    && pProbe->noSkipScan==0
2424    && pProbe->aiRowLogEst[saved_nEq+1]>=42  /* TUNING: Minimum for skip-scan */
2425    && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK
2426   ){
2427     LogEst nIter;
2428     pNew->u.btree.nEq++;
2429     pNew->nSkip++;
2430     pNew->aLTerm[pNew->nLTerm++] = 0;
2431     pNew->wsFlags |= WHERE_SKIPSCAN;
2432     nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1];
2433     pNew->nOut -= nIter;
2434     /* TUNING:  Because uncertainties in the estimates for skip-scan queries,
2435     ** add a 1.375 fudge factor to make skip-scan slightly less likely. */
2436     nIter += 5;
2437     whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul);
2438     pNew->nOut = saved_nOut;
2439     pNew->u.btree.nEq = saved_nEq;
2440     pNew->nSkip = saved_nSkip;
2441     pNew->wsFlags = saved_wsFlags;
2442   }
2443 
2444   return rc;
2445 }
2446 
2447 /*
2448 ** Return True if it is possible that pIndex might be useful in
2449 ** implementing the ORDER BY clause in pBuilder.
2450 **
2451 ** Return False if pBuilder does not contain an ORDER BY clause or
2452 ** if there is no way for pIndex to be useful in implementing that
2453 ** ORDER BY clause.
2454 */
2455 static int indexMightHelpWithOrderBy(
2456   WhereLoopBuilder *pBuilder,
2457   Index *pIndex,
2458   int iCursor
2459 ){
2460   ExprList *pOB;
2461   ExprList *aColExpr;
2462   int ii, jj;
2463 
2464   if( pIndex->bUnordered ) return 0;
2465   if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0;
2466   for(ii=0; ii<pOB->nExpr; ii++){
2467     Expr *pExpr = sqlite3ExprSkipCollate(pOB->a[ii].pExpr);
2468     if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){
2469       if( pExpr->iColumn<0 ) return 1;
2470       for(jj=0; jj<pIndex->nKeyCol; jj++){
2471         if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1;
2472       }
2473     }else if( (aColExpr = pIndex->aColExpr)!=0 ){
2474       for(jj=0; jj<pIndex->nKeyCol; jj++){
2475         if( pIndex->aiColumn[jj]!=XN_EXPR ) continue;
2476         if( sqlite3ExprCompare(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){
2477           return 1;
2478         }
2479       }
2480     }
2481   }
2482   return 0;
2483 }
2484 
2485 /*
2486 ** Return a bitmask where 1s indicate that the corresponding column of
2487 ** the table is used by an index.  Only the first 63 columns are considered.
2488 */
2489 static Bitmask columnsInIndex(Index *pIdx){
2490   Bitmask m = 0;
2491   int j;
2492   for(j=pIdx->nColumn-1; j>=0; j--){
2493     int x = pIdx->aiColumn[j];
2494     if( x>=0 ){
2495       testcase( x==BMS-1 );
2496       testcase( x==BMS-2 );
2497       if( x<BMS-1 ) m |= MASKBIT(x);
2498     }
2499   }
2500   return m;
2501 }
2502 
2503 /* Check to see if a partial index with pPartIndexWhere can be used
2504 ** in the current query.  Return true if it can be and false if not.
2505 */
2506 static int whereUsablePartialIndex(int iTab, WhereClause *pWC, Expr *pWhere){
2507   int i;
2508   WhereTerm *pTerm;
2509   while( pWhere->op==TK_AND ){
2510     if( !whereUsablePartialIndex(iTab,pWC,pWhere->pLeft) ) return 0;
2511     pWhere = pWhere->pRight;
2512   }
2513   for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){
2514     Expr *pExpr = pTerm->pExpr;
2515     if( sqlite3ExprImpliesExpr(pExpr, pWhere, iTab)
2516      && (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab)
2517     ){
2518       return 1;
2519     }
2520   }
2521   return 0;
2522 }
2523 
2524 /*
2525 ** Add all WhereLoop objects for a single table of the join where the table
2526 ** is idenfied by pBuilder->pNew->iTab.  That table is guaranteed to be
2527 ** a b-tree table, not a virtual table.
2528 **
2529 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function
2530 ** are calculated as follows:
2531 **
2532 ** For a full scan, assuming the table (or index) contains nRow rows:
2533 **
2534 **     cost = nRow * 3.0                    // full-table scan
2535 **     cost = nRow * K                      // scan of covering index
2536 **     cost = nRow * (K+3.0)                // scan of non-covering index
2537 **
2538 ** where K is a value between 1.1 and 3.0 set based on the relative
2539 ** estimated average size of the index and table records.
2540 **
2541 ** For an index scan, where nVisit is the number of index rows visited
2542 ** by the scan, and nSeek is the number of seek operations required on
2543 ** the index b-tree:
2544 **
2545 **     cost = nSeek * (log(nRow) + K * nVisit)          // covering index
2546 **     cost = nSeek * (log(nRow) + (K+3.0) * nVisit)    // non-covering index
2547 **
2548 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the
2549 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when
2550 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans.
2551 **
2552 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount
2553 ** of uncertainty.  For this reason, scoring is designed to pick plans that
2554 ** "do the least harm" if the estimates are inaccurate.  For example, a
2555 ** log(nRow) factor is omitted from a non-covering index scan in order to
2556 ** bias the scoring in favor of using an index, since the worst-case
2557 ** performance of using an index is far better than the worst-case performance
2558 ** of a full table scan.
2559 */
2560 static int whereLoopAddBtree(
2561   WhereLoopBuilder *pBuilder, /* WHERE clause information */
2562   Bitmask mPrereq             /* Extra prerequesites for using this table */
2563 ){
2564   WhereInfo *pWInfo;          /* WHERE analysis context */
2565   Index *pProbe;              /* An index we are evaluating */
2566   Index sPk;                  /* A fake index object for the primary key */
2567   LogEst aiRowEstPk[2];       /* The aiRowLogEst[] value for the sPk index */
2568   i16 aiColumnPk = -1;        /* The aColumn[] value for the sPk index */
2569   SrcList *pTabList;          /* The FROM clause */
2570   struct SrcList_item *pSrc;  /* The FROM clause btree term to add */
2571   WhereLoop *pNew;            /* Template WhereLoop object */
2572   int rc = SQLITE_OK;         /* Return code */
2573   int iSortIdx = 1;           /* Index number */
2574   int b;                      /* A boolean value */
2575   LogEst rSize;               /* number of rows in the table */
2576   LogEst rLogSize;            /* Logarithm of the number of rows in the table */
2577   WhereClause *pWC;           /* The parsed WHERE clause */
2578   Table *pTab;                /* Table being queried */
2579 
2580   pNew = pBuilder->pNew;
2581   pWInfo = pBuilder->pWInfo;
2582   pTabList = pWInfo->pTabList;
2583   pSrc = pTabList->a + pNew->iTab;
2584   pTab = pSrc->pTab;
2585   pWC = pBuilder->pWC;
2586   assert( !IsVirtual(pSrc->pTab) );
2587 
2588   if( pSrc->pIBIndex ){
2589     /* An INDEXED BY clause specifies a particular index to use */
2590     pProbe = pSrc->pIBIndex;
2591   }else if( !HasRowid(pTab) ){
2592     pProbe = pTab->pIndex;
2593   }else{
2594     /* There is no INDEXED BY clause.  Create a fake Index object in local
2595     ** variable sPk to represent the rowid primary key index.  Make this
2596     ** fake index the first in a chain of Index objects with all of the real
2597     ** indices to follow */
2598     Index *pFirst;                  /* First of real indices on the table */
2599     memset(&sPk, 0, sizeof(Index));
2600     sPk.nKeyCol = 1;
2601     sPk.nColumn = 1;
2602     sPk.aiColumn = &aiColumnPk;
2603     sPk.aiRowLogEst = aiRowEstPk;
2604     sPk.onError = OE_Replace;
2605     sPk.pTable = pTab;
2606     sPk.szIdxRow = pTab->szTabRow;
2607     aiRowEstPk[0] = pTab->nRowLogEst;
2608     aiRowEstPk[1] = 0;
2609     pFirst = pSrc->pTab->pIndex;
2610     if( pSrc->fg.notIndexed==0 ){
2611       /* The real indices of the table are only considered if the
2612       ** NOT INDEXED qualifier is omitted from the FROM clause */
2613       sPk.pNext = pFirst;
2614     }
2615     pProbe = &sPk;
2616   }
2617   rSize = pTab->nRowLogEst;
2618   rLogSize = estLog(rSize);
2619 
2620 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
2621   /* Automatic indexes */
2622   if( !pBuilder->pOrSet      /* Not part of an OR optimization */
2623    && (pWInfo->wctrlFlags & WHERE_NO_AUTOINDEX)==0
2624    && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0
2625    && pSrc->pIBIndex==0      /* Has no INDEXED BY clause */
2626    && !pSrc->fg.notIndexed   /* Has no NOT INDEXED clause */
2627    && HasRowid(pTab)         /* Not WITHOUT ROWID table. (FIXME: Why not?) */
2628    && !pSrc->fg.isCorrelated /* Not a correlated subquery */
2629    && !pSrc->fg.isRecursive  /* Not a recursive common table expression. */
2630   ){
2631     /* Generate auto-index WhereLoops */
2632     WhereTerm *pTerm;
2633     WhereTerm *pWCEnd = pWC->a + pWC->nTerm;
2634     for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){
2635       if( pTerm->prereqRight & pNew->maskSelf ) continue;
2636       if( termCanDriveIndex(pTerm, pSrc, 0) ){
2637         pNew->u.btree.nEq = 1;
2638         pNew->nSkip = 0;
2639         pNew->u.btree.pIndex = 0;
2640         pNew->nLTerm = 1;
2641         pNew->aLTerm[0] = pTerm;
2642         /* TUNING: One-time cost for computing the automatic index is
2643         ** estimated to be X*N*log2(N) where N is the number of rows in
2644         ** the table being indexed and where X is 7 (LogEst=28) for normal
2645         ** tables or 1.375 (LogEst=4) for views and subqueries.  The value
2646         ** of X is smaller for views and subqueries so that the query planner
2647         ** will be more aggressive about generating automatic indexes for
2648         ** those objects, since there is no opportunity to add schema
2649         ** indexes on subqueries and views. */
2650         pNew->rSetup = rLogSize + rSize + 4;
2651         if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){
2652           pNew->rSetup += 24;
2653         }
2654         ApplyCostMultiplier(pNew->rSetup, pTab->costMult);
2655         /* TUNING: Each index lookup yields 20 rows in the table.  This
2656         ** is more than the usual guess of 10 rows, since we have no way
2657         ** of knowing how selective the index will ultimately be.  It would
2658         ** not be unreasonable to make this value much larger. */
2659         pNew->nOut = 43;  assert( 43==sqlite3LogEst(20) );
2660         pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut);
2661         pNew->wsFlags = WHERE_AUTO_INDEX;
2662         pNew->prereq = mPrereq | pTerm->prereqRight;
2663         rc = whereLoopInsert(pBuilder, pNew);
2664       }
2665     }
2666   }
2667 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */
2668 
2669   /* Loop over all indices
2670   */
2671   for(; rc==SQLITE_OK && pProbe; pProbe=pProbe->pNext, iSortIdx++){
2672     if( pProbe->pPartIdxWhere!=0
2673      && !whereUsablePartialIndex(pSrc->iCursor, pWC, pProbe->pPartIdxWhere) ){
2674       testcase( pNew->iTab!=pSrc->iCursor );  /* See ticket [98d973b8f5] */
2675       continue;  /* Partial index inappropriate for this query */
2676     }
2677     rSize = pProbe->aiRowLogEst[0];
2678     pNew->u.btree.nEq = 0;
2679     pNew->nSkip = 0;
2680     pNew->nLTerm = 0;
2681     pNew->iSortIdx = 0;
2682     pNew->rSetup = 0;
2683     pNew->prereq = mPrereq;
2684     pNew->nOut = rSize;
2685     pNew->u.btree.pIndex = pProbe;
2686     b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor);
2687     /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */
2688     assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 );
2689     if( pProbe->tnum<=0 ){
2690       /* Integer primary key index */
2691       pNew->wsFlags = WHERE_IPK;
2692 
2693       /* Full table scan */
2694       pNew->iSortIdx = b ? iSortIdx : 0;
2695       /* TUNING: Cost of full table scan is (N*3.0). */
2696       pNew->rRun = rSize + 16;
2697       ApplyCostMultiplier(pNew->rRun, pTab->costMult);
2698       whereLoopOutputAdjust(pWC, pNew, rSize);
2699       rc = whereLoopInsert(pBuilder, pNew);
2700       pNew->nOut = rSize;
2701       if( rc ) break;
2702     }else{
2703       Bitmask m;
2704       if( pProbe->isCovering ){
2705         pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED;
2706         m = 0;
2707       }else{
2708         m = pSrc->colUsed & ~columnsInIndex(pProbe);
2709         pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED;
2710       }
2711 
2712       /* Full scan via index */
2713       if( b
2714        || !HasRowid(pTab)
2715        || ( m==0
2716          && pProbe->bUnordered==0
2717          && (pProbe->szIdxRow<pTab->szTabRow)
2718          && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0
2719          && sqlite3GlobalConfig.bUseCis
2720          && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan)
2721           )
2722       ){
2723         pNew->iSortIdx = b ? iSortIdx : 0;
2724 
2725         /* The cost of visiting the index rows is N*K, where K is
2726         ** between 1.1 and 3.0, depending on the relative sizes of the
2727         ** index and table rows. If this is a non-covering index scan,
2728         ** also add the cost of visiting table rows (N*3.0).  */
2729         pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow;
2730         if( m!=0 ){
2731           pNew->rRun = sqlite3LogEstAdd(pNew->rRun, rSize+16);
2732         }
2733         ApplyCostMultiplier(pNew->rRun, pTab->costMult);
2734         whereLoopOutputAdjust(pWC, pNew, rSize);
2735         rc = whereLoopInsert(pBuilder, pNew);
2736         pNew->nOut = rSize;
2737         if( rc ) break;
2738       }
2739     }
2740 
2741     rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0);
2742 #ifdef SQLITE_ENABLE_STAT3_OR_STAT4
2743     sqlite3Stat4ProbeFree(pBuilder->pRec);
2744     pBuilder->nRecValid = 0;
2745     pBuilder->pRec = 0;
2746 #endif
2747 
2748     /* If there was an INDEXED BY clause, then only that one index is
2749     ** considered. */
2750     if( pSrc->pIBIndex ) break;
2751   }
2752   return rc;
2753 }
2754 
2755 #ifndef SQLITE_OMIT_VIRTUALTABLE
2756 
2757 /*
2758 ** Argument pIdxInfo is already populated with all constraints that may
2759 ** be used by the virtual table identified by pBuilder->pNew->iTab. This
2760 ** function marks a subset of those constraints usable, invokes the
2761 ** xBestIndex method and adds the returned plan to pBuilder.
2762 **
2763 ** A constraint is marked usable if:
2764 **
2765 **   * Argument mUsable indicates that its prerequisites are available, and
2766 **
2767 **   * It is not one of the operators specified in the mExclude mask passed
2768 **     as the fourth argument (which in practice is either WO_IN or 0).
2769 **
2770 ** Argument mPrereq is a mask of tables that must be scanned before the
2771 ** virtual table in question. These are added to the plans prerequisites
2772 ** before it is added to pBuilder.
2773 **
2774 ** Output parameter *pbIn is set to true if the plan added to pBuilder
2775 ** uses one or more WO_IN terms, or false otherwise.
2776 */
2777 static int whereLoopAddVirtualOne(
2778   WhereLoopBuilder *pBuilder,
2779   Bitmask mPrereq,                /* Mask of tables that must be used. */
2780   Bitmask mUsable,                /* Mask of usable tables */
2781   u16 mExclude,                   /* Exclude terms using these operators */
2782   sqlite3_index_info *pIdxInfo,   /* Populated object for xBestIndex */
2783   int *pbIn                       /* OUT: True if plan uses an IN(...) op */
2784 ){
2785   WhereClause *pWC = pBuilder->pWC;
2786   struct sqlite3_index_constraint *pIdxCons;
2787   struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage;
2788   int i;
2789   int mxTerm;
2790   int rc = SQLITE_OK;
2791   WhereLoop *pNew = pBuilder->pNew;
2792   Parse *pParse = pBuilder->pWInfo->pParse;
2793   struct SrcList_item *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab];
2794   int nConstraint = pIdxInfo->nConstraint;
2795 
2796   assert( (mUsable & mPrereq)==mPrereq );
2797   *pbIn = 0;
2798   pNew->prereq = mPrereq;
2799 
2800   /* Set the usable flag on the subset of constraints identified by
2801   ** arguments mUsable and mExclude. */
2802   pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
2803   for(i=0; i<nConstraint; i++, pIdxCons++){
2804     WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset];
2805     pIdxCons->usable = 0;
2806     if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight
2807      && (pTerm->eOperator & mExclude)==0
2808     ){
2809       pIdxCons->usable = 1;
2810     }
2811   }
2812 
2813   /* Initialize the output fields of the sqlite3_index_info structure */
2814   memset(pUsage, 0, sizeof(pUsage[0])*nConstraint);
2815   assert( pIdxInfo->needToFreeIdxStr==0 );
2816   pIdxInfo->idxStr = 0;
2817   pIdxInfo->idxNum = 0;
2818   pIdxInfo->orderByConsumed = 0;
2819   pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2;
2820   pIdxInfo->estimatedRows = 25;
2821   pIdxInfo->idxFlags = 0;
2822   pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed;
2823 
2824   /* Invoke the virtual table xBestIndex() method */
2825   rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo);
2826   if( rc ) return rc;
2827 
2828   mxTerm = -1;
2829   assert( pNew->nLSlot>=nConstraint );
2830   for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0;
2831   pNew->u.vtab.omitMask = 0;
2832   pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint;
2833   for(i=0; i<nConstraint; i++, pIdxCons++){
2834     int iTerm;
2835     if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){
2836       WhereTerm *pTerm;
2837       int j = pIdxCons->iTermOffset;
2838       if( iTerm>=nConstraint
2839        || j<0
2840        || j>=pWC->nTerm
2841        || pNew->aLTerm[iTerm]!=0
2842        || pIdxCons->usable==0
2843       ){
2844         rc = SQLITE_ERROR;
2845         sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName);
2846         return rc;
2847       }
2848       testcase( iTerm==nConstraint-1 );
2849       testcase( j==0 );
2850       testcase( j==pWC->nTerm-1 );
2851       pTerm = &pWC->a[j];
2852       pNew->prereq |= pTerm->prereqRight;
2853       assert( iTerm<pNew->nLSlot );
2854       pNew->aLTerm[iTerm] = pTerm;
2855       if( iTerm>mxTerm ) mxTerm = iTerm;
2856       testcase( iTerm==15 );
2857       testcase( iTerm==16 );
2858       if( iTerm<16 && pUsage[i].omit ) pNew->u.vtab.omitMask |= 1<<iTerm;
2859       if( (pTerm->eOperator & WO_IN)!=0 ){
2860         /* A virtual table that is constrained by an IN clause may not
2861         ** consume the ORDER BY clause because (1) the order of IN terms
2862         ** is not necessarily related to the order of output terms and
2863         ** (2) Multiple outputs from a single IN value will not merge
2864         ** together.  */
2865         pIdxInfo->orderByConsumed = 0;
2866         pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE;
2867         *pbIn = 1; assert( (mExclude & WO_IN)==0 );
2868       }
2869     }
2870   }
2871 
2872   pNew->nLTerm = mxTerm+1;
2873   assert( pNew->nLTerm<=pNew->nLSlot );
2874   pNew->u.vtab.idxNum = pIdxInfo->idxNum;
2875   pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr;
2876   pIdxInfo->needToFreeIdxStr = 0;
2877   pNew->u.vtab.idxStr = pIdxInfo->idxStr;
2878   pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ?
2879       pIdxInfo->nOrderBy : 0);
2880   pNew->rSetup = 0;
2881   pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost);
2882   pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows);
2883 
2884   /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated
2885   ** that the scan will visit at most one row. Clear it otherwise. */
2886   if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){
2887     pNew->wsFlags |= WHERE_ONEROW;
2888   }else{
2889     pNew->wsFlags &= ~WHERE_ONEROW;
2890   }
2891   rc = whereLoopInsert(pBuilder, pNew);
2892   if( pNew->u.vtab.needFree ){
2893     sqlite3_free(pNew->u.vtab.idxStr);
2894     pNew->u.vtab.needFree = 0;
2895   }
2896   WHERETRACE(0xffff, ("  bIn=%d prereqIn=%04llx prereqOut=%04llx\n",
2897                       *pbIn, (sqlite3_uint64)mPrereq,
2898                       (sqlite3_uint64)(pNew->prereq & ~mPrereq)));
2899 
2900   return rc;
2901 }
2902 
2903 
2904 /*
2905 ** Add all WhereLoop objects for a table of the join identified by
2906 ** pBuilder->pNew->iTab.  That table is guaranteed to be a virtual table.
2907 **
2908 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and
2909 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause
2910 ** entries that occur before the virtual table in the FROM clause and are
2911 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the
2912 ** mUnusable mask contains all FROM clause entries that occur after the
2913 ** virtual table and are separated from it by at least one LEFT or
2914 ** CROSS JOIN.
2915 **
2916 ** For example, if the query were:
2917 **
2918 **   ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6;
2919 **
2920 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6).
2921 **
2922 ** All the tables in mPrereq must be scanned before the current virtual
2923 ** table. So any terms for which all prerequisites are satisfied by
2924 ** mPrereq may be specified as "usable" in all calls to xBestIndex.
2925 ** Conversely, all tables in mUnusable must be scanned after the current
2926 ** virtual table, so any terms for which the prerequisites overlap with
2927 ** mUnusable should always be configured as "not-usable" for xBestIndex.
2928 */
2929 static int whereLoopAddVirtual(
2930   WhereLoopBuilder *pBuilder,  /* WHERE clause information */
2931   Bitmask mPrereq,             /* Tables that must be scanned before this one */
2932   Bitmask mUnusable            /* Tables that must be scanned after this one */
2933 ){
2934   int rc = SQLITE_OK;          /* Return code */
2935   WhereInfo *pWInfo;           /* WHERE analysis context */
2936   Parse *pParse;               /* The parsing context */
2937   WhereClause *pWC;            /* The WHERE clause */
2938   struct SrcList_item *pSrc;   /* The FROM clause term to search */
2939   sqlite3_index_info *p;       /* Object to pass to xBestIndex() */
2940   int nConstraint;             /* Number of constraints in p */
2941   int bIn;                     /* True if plan uses IN(...) operator */
2942   WhereLoop *pNew;
2943   Bitmask mBest;               /* Tables used by best possible plan */
2944 
2945   assert( (mPrereq & mUnusable)==0 );
2946   pWInfo = pBuilder->pWInfo;
2947   pParse = pWInfo->pParse;
2948   pWC = pBuilder->pWC;
2949   pNew = pBuilder->pNew;
2950   pSrc = &pWInfo->pTabList->a[pNew->iTab];
2951   assert( IsVirtual(pSrc->pTab) );
2952   p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy);
2953   if( p==0 ) return SQLITE_NOMEM_BKPT;
2954   pNew->rSetup = 0;
2955   pNew->wsFlags = WHERE_VIRTUALTABLE;
2956   pNew->nLTerm = 0;
2957   pNew->u.vtab.needFree = 0;
2958   nConstraint = p->nConstraint;
2959   if( whereLoopResize(pParse->db, pNew, nConstraint) ){
2960     sqlite3DbFree(pParse->db, p);
2961     return SQLITE_NOMEM_BKPT;
2962   }
2963 
2964   /* First call xBestIndex() with all constraints usable. */
2965   WHERETRACE(0x40, ("  VirtualOne: all usable\n"));
2966   rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, &bIn);
2967 
2968   /* If the call to xBestIndex() with all terms enabled produced a plan
2969   ** that does not require any source tables (IOW: a plan with mBest==0),
2970   ** then there is no point in making any further calls to xBestIndex()
2971   ** since they will all return the same result (if the xBestIndex()
2972   ** implementation is sane). */
2973   if( rc==SQLITE_OK && (mBest = (pNew->prereq & ~mPrereq))!=0 ){
2974     int seenZero = 0;             /* True if a plan with no prereqs seen */
2975     int seenZeroNoIN = 0;         /* Plan with no prereqs and no IN(...) seen */
2976     Bitmask mPrev = 0;
2977     Bitmask mBestNoIn = 0;
2978 
2979     /* If the plan produced by the earlier call uses an IN(...) term, call
2980     ** xBestIndex again, this time with IN(...) terms disabled. */
2981     if( bIn ){
2982       WHERETRACE(0x40, ("  VirtualOne: all usable w/o IN\n"));
2983       rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, WO_IN, p, &bIn);
2984       assert( bIn==0 );
2985       mBestNoIn = pNew->prereq & ~mPrereq;
2986       if( mBestNoIn==0 ){
2987         seenZero = 1;
2988         seenZeroNoIN = 1;
2989       }
2990     }
2991 
2992     /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq)
2993     ** in the set of terms that apply to the current virtual table.  */
2994     while( rc==SQLITE_OK ){
2995       int i;
2996       Bitmask mNext = ALLBITS;
2997       assert( mNext>0 );
2998       for(i=0; i<nConstraint; i++){
2999         Bitmask mThis = (
3000             pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq
3001         );
3002         if( mThis>mPrev && mThis<mNext ) mNext = mThis;
3003       }
3004       mPrev = mNext;
3005       if( mNext==ALLBITS ) break;
3006       if( mNext==mBest || mNext==mBestNoIn ) continue;
3007       WHERETRACE(0x40, ("  VirtualOne: mPrev=%04llx mNext=%04llx\n",
3008                        (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext));
3009       rc = whereLoopAddVirtualOne(pBuilder, mPrereq, mNext|mPrereq, 0, p, &bIn);
3010       if( pNew->prereq==mPrereq ){
3011         seenZero = 1;
3012         if( bIn==0 ) seenZeroNoIN = 1;
3013       }
3014     }
3015 
3016     /* If the calls to xBestIndex() in the above loop did not find a plan
3017     ** that requires no source tables at all (i.e. one guaranteed to be
3018     ** usable), make a call here with all source tables disabled */
3019     if( rc==SQLITE_OK && seenZero==0 ){
3020       WHERETRACE(0x40, ("  VirtualOne: all disabled\n"));
3021       rc = whereLoopAddVirtualOne(pBuilder, mPrereq, mPrereq, 0, p, &bIn);
3022       if( bIn==0 ) seenZeroNoIN = 1;
3023     }
3024 
3025     /* If the calls to xBestIndex() have so far failed to find a plan
3026     ** that requires no source tables at all and does not use an IN(...)
3027     ** operator, make a final call to obtain one here.  */
3028     if( rc==SQLITE_OK && seenZeroNoIN==0 ){
3029       WHERETRACE(0x40, ("  VirtualOne: all disabled and w/o IN\n"));
3030       rc = whereLoopAddVirtualOne(pBuilder, mPrereq, mPrereq, WO_IN, p, &bIn);
3031     }
3032   }
3033 
3034   if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr);
3035   sqlite3DbFree(pParse->db, p);
3036   return rc;
3037 }
3038 #endif /* SQLITE_OMIT_VIRTUALTABLE */
3039 
3040 /*
3041 ** Add WhereLoop entries to handle OR terms.  This works for either
3042 ** btrees or virtual tables.
3043 */
3044 static int whereLoopAddOr(
3045   WhereLoopBuilder *pBuilder,
3046   Bitmask mPrereq,
3047   Bitmask mUnusable
3048 ){
3049   WhereInfo *pWInfo = pBuilder->pWInfo;
3050   WhereClause *pWC;
3051   WhereLoop *pNew;
3052   WhereTerm *pTerm, *pWCEnd;
3053   int rc = SQLITE_OK;
3054   int iCur;
3055   WhereClause tempWC;
3056   WhereLoopBuilder sSubBuild;
3057   WhereOrSet sSum, sCur;
3058   struct SrcList_item *pItem;
3059 
3060   pWC = pBuilder->pWC;
3061   pWCEnd = pWC->a + pWC->nTerm;
3062   pNew = pBuilder->pNew;
3063   memset(&sSum, 0, sizeof(sSum));
3064   pItem = pWInfo->pTabList->a + pNew->iTab;
3065   iCur = pItem->iCursor;
3066 
3067   for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){
3068     if( (pTerm->eOperator & WO_OR)!=0
3069      && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0
3070     ){
3071       WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc;
3072       WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm];
3073       WhereTerm *pOrTerm;
3074       int once = 1;
3075       int i, j;
3076 
3077       sSubBuild = *pBuilder;
3078       sSubBuild.pOrderBy = 0;
3079       sSubBuild.pOrSet = &sCur;
3080 
3081       WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm));
3082       for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){
3083         if( (pOrTerm->eOperator & WO_AND)!=0 ){
3084           sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc;
3085         }else if( pOrTerm->leftCursor==iCur ){
3086           tempWC.pWInfo = pWC->pWInfo;
3087           tempWC.pOuter = pWC;
3088           tempWC.op = TK_AND;
3089           tempWC.nTerm = 1;
3090           tempWC.a = pOrTerm;
3091           sSubBuild.pWC = &tempWC;
3092         }else{
3093           continue;
3094         }
3095         sCur.n = 0;
3096 #ifdef WHERETRACE_ENABLED
3097         WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n",
3098                    (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm));
3099         if( sqlite3WhereTrace & 0x400 ){
3100           for(i=0; i<sSubBuild.pWC->nTerm; i++){
3101             whereTermPrint(&sSubBuild.pWC->a[i], i);
3102           }
3103         }
3104 #endif
3105 #ifndef SQLITE_OMIT_VIRTUALTABLE
3106         if( IsVirtual(pItem->pTab) ){
3107           rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable);
3108         }else
3109 #endif
3110         {
3111           rc = whereLoopAddBtree(&sSubBuild, mPrereq);
3112         }
3113         if( rc==SQLITE_OK ){
3114           rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable);
3115         }
3116         assert( rc==SQLITE_OK || sCur.n==0 );
3117         if( sCur.n==0 ){
3118           sSum.n = 0;
3119           break;
3120         }else if( once ){
3121           whereOrMove(&sSum, &sCur);
3122           once = 0;
3123         }else{
3124           WhereOrSet sPrev;
3125           whereOrMove(&sPrev, &sSum);
3126           sSum.n = 0;
3127           for(i=0; i<sPrev.n; i++){
3128             for(j=0; j<sCur.n; j++){
3129               whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq,
3130                             sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun),
3131                             sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut));
3132             }
3133           }
3134         }
3135       }
3136       pNew->nLTerm = 1;
3137       pNew->aLTerm[0] = pTerm;
3138       pNew->wsFlags = WHERE_MULTI_OR;
3139       pNew->rSetup = 0;
3140       pNew->iSortIdx = 0;
3141       memset(&pNew->u, 0, sizeof(pNew->u));
3142       for(i=0; rc==SQLITE_OK && i<sSum.n; i++){
3143         /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs
3144         ** of all sub-scans required by the OR-scan. However, due to rounding
3145         ** errors, it may be that the cost of the OR-scan is equal to its
3146         ** most expensive sub-scan. Add the smallest possible penalty
3147         ** (equivalent to multiplying the cost by 1.07) to ensure that
3148         ** this does not happen. Otherwise, for WHERE clauses such as the
3149         ** following where there is an index on "y":
3150         **
3151         **     WHERE likelihood(x=?, 0.99) OR y=?
3152         **
3153         ** the planner may elect to "OR" together a full-table scan and an
3154         ** index lookup. And other similarly odd results.  */
3155         pNew->rRun = sSum.a[i].rRun + 1;
3156         pNew->nOut = sSum.a[i].nOut;
3157         pNew->prereq = sSum.a[i].prereq;
3158         rc = whereLoopInsert(pBuilder, pNew);
3159       }
3160       WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm));
3161     }
3162   }
3163   return rc;
3164 }
3165 
3166 /*
3167 ** Add all WhereLoop objects for all tables
3168 */
3169 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){
3170   WhereInfo *pWInfo = pBuilder->pWInfo;
3171   Bitmask mPrereq = 0;
3172   Bitmask mPrior = 0;
3173   int iTab;
3174   SrcList *pTabList = pWInfo->pTabList;
3175   struct SrcList_item *pItem;
3176   struct SrcList_item *pEnd = &pTabList->a[pWInfo->nLevel];
3177   sqlite3 *db = pWInfo->pParse->db;
3178   int rc = SQLITE_OK;
3179   WhereLoop *pNew;
3180   u8 priorJointype = 0;
3181 
3182   /* Loop over the tables in the join, from left to right */
3183   pNew = pBuilder->pNew;
3184   whereLoopInit(pNew);
3185   for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){
3186     Bitmask mUnusable = 0;
3187     pNew->iTab = iTab;
3188     pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor);
3189     if( ((pItem->fg.jointype|priorJointype) & (JT_LEFT|JT_CROSS))!=0 ){
3190       /* This condition is true when pItem is the FROM clause term on the
3191       ** right-hand-side of a LEFT or CROSS JOIN.  */
3192       mPrereq = mPrior;
3193     }
3194     priorJointype = pItem->fg.jointype;
3195     if( IsVirtual(pItem->pTab) ){
3196       struct SrcList_item *p;
3197       for(p=&pItem[1]; p<pEnd; p++){
3198         if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){
3199           mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor);
3200         }
3201       }
3202       rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable);
3203     }else{
3204       rc = whereLoopAddBtree(pBuilder, mPrereq);
3205     }
3206     if( rc==SQLITE_OK ){
3207       rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable);
3208     }
3209     mPrior |= pNew->maskSelf;
3210     if( rc || db->mallocFailed ) break;
3211   }
3212 
3213   whereLoopClear(db, pNew);
3214   return rc;
3215 }
3216 
3217 /*
3218 ** Examine a WherePath (with the addition of the extra WhereLoop of the 5th
3219 ** parameters) to see if it outputs rows in the requested ORDER BY
3220 ** (or GROUP BY) without requiring a separate sort operation.  Return N:
3221 **
3222 **   N>0:   N terms of the ORDER BY clause are satisfied
3223 **   N==0:  No terms of the ORDER BY clause are satisfied
3224 **   N<0:   Unknown yet how many terms of ORDER BY might be satisfied.
3225 **
3226 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as
3227 ** strict.  With GROUP BY and DISTINCT the only requirement is that
3228 ** equivalent rows appear immediately adjacent to one another.  GROUP BY
3229 ** and DISTINCT do not require rows to appear in any particular order as long
3230 ** as equivalent rows are grouped together.  Thus for GROUP BY and DISTINCT
3231 ** the pOrderBy terms can be matched in any order.  With ORDER BY, the
3232 ** pOrderBy terms must be matched in strict left-to-right order.
3233 */
3234 static i8 wherePathSatisfiesOrderBy(
3235   WhereInfo *pWInfo,    /* The WHERE clause */
3236   ExprList *pOrderBy,   /* ORDER BY or GROUP BY or DISTINCT clause to check */
3237   WherePath *pPath,     /* The WherePath to check */
3238   u16 wctrlFlags,       /* Might contain WHERE_GROUPBY or WHERE_DISTINCTBY */
3239   u16 nLoop,            /* Number of entries in pPath->aLoop[] */
3240   WhereLoop *pLast,     /* Add this WhereLoop to the end of pPath->aLoop[] */
3241   Bitmask *pRevMask     /* OUT: Mask of WhereLoops to run in reverse order */
3242 ){
3243   u8 revSet;            /* True if rev is known */
3244   u8 rev;               /* Composite sort order */
3245   u8 revIdx;            /* Index sort order */
3246   u8 isOrderDistinct;   /* All prior WhereLoops are order-distinct */
3247   u8 distinctColumns;   /* True if the loop has UNIQUE NOT NULL columns */
3248   u8 isMatch;           /* iColumn matches a term of the ORDER BY clause */
3249   u16 nKeyCol;          /* Number of key columns in pIndex */
3250   u16 nColumn;          /* Total number of ordered columns in the index */
3251   u16 nOrderBy;         /* Number terms in the ORDER BY clause */
3252   int iLoop;            /* Index of WhereLoop in pPath being processed */
3253   int i, j;             /* Loop counters */
3254   int iCur;             /* Cursor number for current WhereLoop */
3255   int iColumn;          /* A column number within table iCur */
3256   WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */
3257   WhereTerm *pTerm;     /* A single term of the WHERE clause */
3258   Expr *pOBExpr;        /* An expression from the ORDER BY clause */
3259   CollSeq *pColl;       /* COLLATE function from an ORDER BY clause term */
3260   Index *pIndex;        /* The index associated with pLoop */
3261   sqlite3 *db = pWInfo->pParse->db;  /* Database connection */
3262   Bitmask obSat = 0;    /* Mask of ORDER BY terms satisfied so far */
3263   Bitmask obDone;       /* Mask of all ORDER BY terms */
3264   Bitmask orderDistinctMask;  /* Mask of all well-ordered loops */
3265   Bitmask ready;              /* Mask of inner loops */
3266 
3267   /*
3268   ** We say the WhereLoop is "one-row" if it generates no more than one
3269   ** row of output.  A WhereLoop is one-row if all of the following are true:
3270   **  (a) All index columns match with WHERE_COLUMN_EQ.
3271   **  (b) The index is unique
3272   ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row.
3273   ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags.
3274   **
3275   ** We say the WhereLoop is "order-distinct" if the set of columns from
3276   ** that WhereLoop that are in the ORDER BY clause are different for every
3277   ** row of the WhereLoop.  Every one-row WhereLoop is automatically
3278   ** order-distinct.   A WhereLoop that has no columns in the ORDER BY clause
3279   ** is not order-distinct. To be order-distinct is not quite the same as being
3280   ** UNIQUE since a UNIQUE column or index can have multiple rows that
3281   ** are NULL and NULL values are equivalent for the purpose of order-distinct.
3282   ** To be order-distinct, the columns must be UNIQUE and NOT NULL.
3283   **
3284   ** The rowid for a table is always UNIQUE and NOT NULL so whenever the
3285   ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is
3286   ** automatically order-distinct.
3287   */
3288 
3289   assert( pOrderBy!=0 );
3290   if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0;
3291 
3292   nOrderBy = pOrderBy->nExpr;
3293   testcase( nOrderBy==BMS-1 );
3294   if( nOrderBy>BMS-1 ) return 0;  /* Cannot optimize overly large ORDER BYs */
3295   isOrderDistinct = 1;
3296   obDone = MASKBIT(nOrderBy)-1;
3297   orderDistinctMask = 0;
3298   ready = 0;
3299   for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){
3300     if( iLoop>0 ) ready |= pLoop->maskSelf;
3301     pLoop = iLoop<nLoop ? pPath->aLoop[iLoop] : pLast;
3302     if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){
3303       if( pLoop->u.vtab.isOrdered ) obSat = obDone;
3304       break;
3305     }
3306     iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor;
3307 
3308     /* Mark off any ORDER BY term X that is a column in the table of
3309     ** the current loop for which there is term in the WHERE
3310     ** clause of the form X IS NULL or X=? that reference only outer
3311     ** loops.
3312     */
3313     for(i=0; i<nOrderBy; i++){
3314       if( MASKBIT(i) & obSat ) continue;
3315       pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr);
3316       if( pOBExpr->op!=TK_COLUMN ) continue;
3317       if( pOBExpr->iTable!=iCur ) continue;
3318       pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn,
3319                        ~ready, WO_EQ|WO_ISNULL|WO_IS, 0);
3320       if( pTerm==0 ) continue;
3321       if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){
3322         const char *z1, *z2;
3323         pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr);
3324         if( !pColl ) pColl = db->pDfltColl;
3325         z1 = pColl->zName;
3326         pColl = sqlite3ExprCollSeq(pWInfo->pParse, pTerm->pExpr);
3327         if( !pColl ) pColl = db->pDfltColl;
3328         z2 = pColl->zName;
3329         if( sqlite3StrICmp(z1, z2)!=0 ) continue;
3330         testcase( pTerm->pExpr->op==TK_IS );
3331       }
3332       obSat |= MASKBIT(i);
3333     }
3334 
3335     if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){
3336       if( pLoop->wsFlags & WHERE_IPK ){
3337         pIndex = 0;
3338         nKeyCol = 0;
3339         nColumn = 1;
3340       }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){
3341         return 0;
3342       }else{
3343         nKeyCol = pIndex->nKeyCol;
3344         nColumn = pIndex->nColumn;
3345         assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) );
3346         assert( pIndex->aiColumn[nColumn-1]==XN_ROWID
3347                           || !HasRowid(pIndex->pTable));
3348         isOrderDistinct = IsUniqueIndex(pIndex);
3349       }
3350 
3351       /* Loop through all columns of the index and deal with the ones
3352       ** that are not constrained by == or IN.
3353       */
3354       rev = revSet = 0;
3355       distinctColumns = 0;
3356       for(j=0; j<nColumn; j++){
3357         u8 bOnce;   /* True to run the ORDER BY search loop */
3358 
3359         /* Skip over == and IS NULL terms */
3360         if( j<pLoop->u.btree.nEq
3361          && pLoop->nSkip==0
3362          && ((i = pLoop->aLTerm[j]->eOperator) & (WO_EQ|WO_ISNULL|WO_IS))!=0
3363         ){
3364           if( i & WO_ISNULL ){
3365             testcase( isOrderDistinct );
3366             isOrderDistinct = 0;
3367           }
3368           continue;
3369         }
3370 
3371         /* Get the column number in the table (iColumn) and sort order
3372         ** (revIdx) for the j-th column of the index.
3373         */
3374         if( pIndex ){
3375           iColumn = pIndex->aiColumn[j];
3376           revIdx = pIndex->aSortOrder[j];
3377           if( iColumn==pIndex->pTable->iPKey ) iColumn = -1;
3378         }else{
3379           iColumn = XN_ROWID;
3380           revIdx = 0;
3381         }
3382 
3383         /* An unconstrained column that might be NULL means that this
3384         ** WhereLoop is not well-ordered
3385         */
3386         if( isOrderDistinct
3387          && iColumn>=0
3388          && j>=pLoop->u.btree.nEq
3389          && pIndex->pTable->aCol[iColumn].notNull==0
3390         ){
3391           isOrderDistinct = 0;
3392         }
3393 
3394         /* Find the ORDER BY term that corresponds to the j-th column
3395         ** of the index and mark that ORDER BY term off
3396         */
3397         bOnce = 1;
3398         isMatch = 0;
3399         for(i=0; bOnce && i<nOrderBy; i++){
3400           if( MASKBIT(i) & obSat ) continue;
3401           pOBExpr = sqlite3ExprSkipCollate(pOrderBy->a[i].pExpr);
3402           testcase( wctrlFlags & WHERE_GROUPBY );
3403           testcase( wctrlFlags & WHERE_DISTINCTBY );
3404           if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0;
3405           if( iColumn>=(-1) ){
3406             if( pOBExpr->op!=TK_COLUMN ) continue;
3407             if( pOBExpr->iTable!=iCur ) continue;
3408             if( pOBExpr->iColumn!=iColumn ) continue;
3409           }else{
3410             if( sqlite3ExprCompare(pOBExpr,pIndex->aColExpr->a[j].pExpr,iCur) ){
3411               continue;
3412             }
3413           }
3414           if( iColumn>=0 ){
3415             pColl = sqlite3ExprCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr);
3416             if( !pColl ) pColl = db->pDfltColl;
3417             if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue;
3418           }
3419           isMatch = 1;
3420           break;
3421         }
3422         if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){
3423           /* Make sure the sort order is compatible in an ORDER BY clause.
3424           ** Sort order is irrelevant for a GROUP BY clause. */
3425           if( revSet ){
3426             if( (rev ^ revIdx)!=pOrderBy->a[i].sortOrder ) isMatch = 0;
3427           }else{
3428             rev = revIdx ^ pOrderBy->a[i].sortOrder;
3429             if( rev ) *pRevMask |= MASKBIT(iLoop);
3430             revSet = 1;
3431           }
3432         }
3433         if( isMatch ){
3434           if( iColumn<0 ){
3435             testcase( distinctColumns==0 );
3436             distinctColumns = 1;
3437           }
3438           obSat |= MASKBIT(i);
3439         }else{
3440           /* No match found */
3441           if( j==0 || j<nKeyCol ){
3442             testcase( isOrderDistinct!=0 );
3443             isOrderDistinct = 0;
3444           }
3445           break;
3446         }
3447       } /* end Loop over all index columns */
3448       if( distinctColumns ){
3449         testcase( isOrderDistinct==0 );
3450         isOrderDistinct = 1;
3451       }
3452     } /* end-if not one-row */
3453 
3454     /* Mark off any other ORDER BY terms that reference pLoop */
3455     if( isOrderDistinct ){
3456       orderDistinctMask |= pLoop->maskSelf;
3457       for(i=0; i<nOrderBy; i++){
3458         Expr *p;
3459         Bitmask mTerm;
3460         if( MASKBIT(i) & obSat ) continue;
3461         p = pOrderBy->a[i].pExpr;
3462         mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p);
3463         if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue;
3464         if( (mTerm&~orderDistinctMask)==0 ){
3465           obSat |= MASKBIT(i);
3466         }
3467       }
3468     }
3469   } /* End the loop over all WhereLoops from outer-most down to inner-most */
3470   if( obSat==obDone ) return (i8)nOrderBy;
3471   if( !isOrderDistinct ){
3472     for(i=nOrderBy-1; i>0; i--){
3473       Bitmask m = MASKBIT(i) - 1;
3474       if( (obSat&m)==m ) return i;
3475     }
3476     return 0;
3477   }
3478   return -1;
3479 }
3480 
3481 
3482 /*
3483 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(),
3484 ** the planner assumes that the specified pOrderBy list is actually a GROUP
3485 ** BY clause - and so any order that groups rows as required satisfies the
3486 ** request.
3487 **
3488 ** Normally, in this case it is not possible for the caller to determine
3489 ** whether or not the rows are really being delivered in sorted order, or
3490 ** just in some other order that provides the required grouping. However,
3491 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then
3492 ** this function may be called on the returned WhereInfo object. It returns
3493 ** true if the rows really will be sorted in the specified order, or false
3494 ** otherwise.
3495 **
3496 ** For example, assuming:
3497 **
3498 **   CREATE INDEX i1 ON t1(x, Y);
3499 **
3500 ** then
3501 **
3502 **   SELECT * FROM t1 GROUP BY x,y ORDER BY x,y;   -- IsSorted()==1
3503 **   SELECT * FROM t1 GROUP BY y,x ORDER BY y,x;   -- IsSorted()==0
3504 */
3505 int sqlite3WhereIsSorted(WhereInfo *pWInfo){
3506   assert( pWInfo->wctrlFlags & WHERE_GROUPBY );
3507   assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP );
3508   return pWInfo->sorted;
3509 }
3510 
3511 #ifdef WHERETRACE_ENABLED
3512 /* For debugging use only: */
3513 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){
3514   static char zName[65];
3515   int i;
3516   for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; }
3517   if( pLast ) zName[i++] = pLast->cId;
3518   zName[i] = 0;
3519   return zName;
3520 }
3521 #endif
3522 
3523 /*
3524 ** Return the cost of sorting nRow rows, assuming that the keys have
3525 ** nOrderby columns and that the first nSorted columns are already in
3526 ** order.
3527 */
3528 static LogEst whereSortingCost(
3529   WhereInfo *pWInfo,
3530   LogEst nRow,
3531   int nOrderBy,
3532   int nSorted
3533 ){
3534   /* TUNING: Estimated cost of a full external sort, where N is
3535   ** the number of rows to sort is:
3536   **
3537   **   cost = (3.0 * N * log(N)).
3538   **
3539   ** Or, if the order-by clause has X terms but only the last Y
3540   ** terms are out of order, then block-sorting will reduce the
3541   ** sorting cost to:
3542   **
3543   **   cost = (3.0 * N * log(N)) * (Y/X)
3544   **
3545   ** The (Y/X) term is implemented using stack variable rScale
3546   ** below.  */
3547   LogEst rScale, rSortCost;
3548   assert( nOrderBy>0 && 66==sqlite3LogEst(100) );
3549   rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66;
3550   rSortCost = nRow + rScale + 16;
3551 
3552   /* Multiple by log(M) where M is the number of output rows.
3553   ** Use the LIMIT for M if it is smaller */
3554   if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){
3555     nRow = pWInfo->iLimit;
3556   }
3557   rSortCost += estLog(nRow);
3558   return rSortCost;
3559 }
3560 
3561 /*
3562 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine
3563 ** attempts to find the lowest cost path that visits each WhereLoop
3564 ** once.  This path is then loaded into the pWInfo->a[].pWLoop fields.
3565 **
3566 ** Assume that the total number of output rows that will need to be sorted
3567 ** will be nRowEst (in the 10*log2 representation).  Or, ignore sorting
3568 ** costs if nRowEst==0.
3569 **
3570 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation
3571 ** error occurs.
3572 */
3573 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){
3574   int mxChoice;             /* Maximum number of simultaneous paths tracked */
3575   int nLoop;                /* Number of terms in the join */
3576   Parse *pParse;            /* Parsing context */
3577   sqlite3 *db;              /* The database connection */
3578   int iLoop;                /* Loop counter over the terms of the join */
3579   int ii, jj;               /* Loop counters */
3580   int mxI = 0;              /* Index of next entry to replace */
3581   int nOrderBy;             /* Number of ORDER BY clause terms */
3582   LogEst mxCost = 0;        /* Maximum cost of a set of paths */
3583   LogEst mxUnsorted = 0;    /* Maximum unsorted cost of a set of path */
3584   int nTo, nFrom;           /* Number of valid entries in aTo[] and aFrom[] */
3585   WherePath *aFrom;         /* All nFrom paths at the previous level */
3586   WherePath *aTo;           /* The nTo best paths at the current level */
3587   WherePath *pFrom;         /* An element of aFrom[] that we are working on */
3588   WherePath *pTo;           /* An element of aTo[] that we are working on */
3589   WhereLoop *pWLoop;        /* One of the WhereLoop objects */
3590   WhereLoop **pX;           /* Used to divy up the pSpace memory */
3591   LogEst *aSortCost = 0;    /* Sorting and partial sorting costs */
3592   char *pSpace;             /* Temporary memory used by this routine */
3593   int nSpace;               /* Bytes of space allocated at pSpace */
3594 
3595   pParse = pWInfo->pParse;
3596   db = pParse->db;
3597   nLoop = pWInfo->nLevel;
3598   /* TUNING: For simple queries, only the best path is tracked.
3599   ** For 2-way joins, the 5 best paths are followed.
3600   ** For joins of 3 or more tables, track the 10 best paths */
3601   mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10);
3602   assert( nLoop<=pWInfo->pTabList->nSrc );
3603   WHERETRACE(0x002, ("---- begin solver.  (nRowEst=%d)\n", nRowEst));
3604 
3605   /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this
3606   ** case the purpose of this call is to estimate the number of rows returned
3607   ** by the overall query. Once this estimate has been obtained, the caller
3608   ** will invoke this function a second time, passing the estimate as the
3609   ** nRowEst parameter.  */
3610   if( pWInfo->pOrderBy==0 || nRowEst==0 ){
3611     nOrderBy = 0;
3612   }else{
3613     nOrderBy = pWInfo->pOrderBy->nExpr;
3614   }
3615 
3616   /* Allocate and initialize space for aTo, aFrom and aSortCost[] */
3617   nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2;
3618   nSpace += sizeof(LogEst) * nOrderBy;
3619   pSpace = sqlite3DbMallocRawNN(db, nSpace);
3620   if( pSpace==0 ) return SQLITE_NOMEM_BKPT;
3621   aTo = (WherePath*)pSpace;
3622   aFrom = aTo+mxChoice;
3623   memset(aFrom, 0, sizeof(aFrom[0]));
3624   pX = (WhereLoop**)(aFrom+mxChoice);
3625   for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){
3626     pFrom->aLoop = pX;
3627   }
3628   if( nOrderBy ){
3629     /* If there is an ORDER BY clause and it is not being ignored, set up
3630     ** space for the aSortCost[] array. Each element of the aSortCost array
3631     ** is either zero - meaning it has not yet been initialized - or the
3632     ** cost of sorting nRowEst rows of data where the first X terms of
3633     ** the ORDER BY clause are already in order, where X is the array
3634     ** index.  */
3635     aSortCost = (LogEst*)pX;
3636     memset(aSortCost, 0, sizeof(LogEst) * nOrderBy);
3637   }
3638   assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] );
3639   assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX );
3640 
3641   /* Seed the search with a single WherePath containing zero WhereLoops.
3642   **
3643   ** TUNING: Do not let the number of iterations go above 28.  If the cost
3644   ** of computing an automatic index is not paid back within the first 28
3645   ** rows, then do not use the automatic index. */
3646   aFrom[0].nRow = MIN(pParse->nQueryLoop, 48);  assert( 48==sqlite3LogEst(28) );
3647   nFrom = 1;
3648   assert( aFrom[0].isOrdered==0 );
3649   if( nOrderBy ){
3650     /* If nLoop is zero, then there are no FROM terms in the query. Since
3651     ** in this case the query may return a maximum of one row, the results
3652     ** are already in the requested order. Set isOrdered to nOrderBy to
3653     ** indicate this. Or, if nLoop is greater than zero, set isOrdered to
3654     ** -1, indicating that the result set may or may not be ordered,
3655     ** depending on the loops added to the current plan.  */
3656     aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy;
3657   }
3658 
3659   /* Compute successively longer WherePaths using the previous generation
3660   ** of WherePaths as the basis for the next.  Keep track of the mxChoice
3661   ** best paths at each generation */
3662   for(iLoop=0; iLoop<nLoop; iLoop++){
3663     nTo = 0;
3664     for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){
3665       for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){
3666         LogEst nOut;                      /* Rows visited by (pFrom+pWLoop) */
3667         LogEst rCost;                     /* Cost of path (pFrom+pWLoop) */
3668         LogEst rUnsorted;                 /* Unsorted cost of (pFrom+pWLoop) */
3669         i8 isOrdered = pFrom->isOrdered;  /* isOrdered for (pFrom+pWLoop) */
3670         Bitmask maskNew;                  /* Mask of src visited by (..) */
3671         Bitmask revMask = 0;              /* Mask of rev-order loops for (..) */
3672 
3673         if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue;
3674         if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue;
3675         if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<10 ){
3676           /* Do not use an automatic index if the this loop is expected
3677           ** to run less than 2 times. */
3678           assert( 10==sqlite3LogEst(2) );
3679           continue;
3680         }
3681         /* At this point, pWLoop is a candidate to be the next loop.
3682         ** Compute its cost */
3683         rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow);
3684         rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted);
3685         nOut = pFrom->nRow + pWLoop->nOut;
3686         maskNew = pFrom->maskLoop | pWLoop->maskSelf;
3687         if( isOrdered<0 ){
3688           isOrdered = wherePathSatisfiesOrderBy(pWInfo,
3689                        pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags,
3690                        iLoop, pWLoop, &revMask);
3691         }else{
3692           revMask = pFrom->revLoop;
3693         }
3694         if( isOrdered>=0 && isOrdered<nOrderBy ){
3695           if( aSortCost[isOrdered]==0 ){
3696             aSortCost[isOrdered] = whereSortingCost(
3697                 pWInfo, nRowEst, nOrderBy, isOrdered
3698             );
3699           }
3700           rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]);
3701 
3702           WHERETRACE(0x002,
3703               ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n",
3704                aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy,
3705                rUnsorted, rCost));
3706         }else{
3707           rCost = rUnsorted;
3708         }
3709 
3710         /* Check to see if pWLoop should be added to the set of
3711         ** mxChoice best-so-far paths.
3712         **
3713         ** First look for an existing path among best-so-far paths
3714         ** that covers the same set of loops and has the same isOrdered
3715         ** setting as the current path candidate.
3716         **
3717         ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent
3718         ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range
3719         ** of legal values for isOrdered, -1..64.
3720         */
3721         for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){
3722           if( pTo->maskLoop==maskNew
3723            && ((pTo->isOrdered^isOrdered)&0x80)==0
3724           ){
3725             testcase( jj==nTo-1 );
3726             break;
3727           }
3728         }
3729         if( jj>=nTo ){
3730           /* None of the existing best-so-far paths match the candidate. */
3731           if( nTo>=mxChoice
3732            && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted))
3733           ){
3734             /* The current candidate is no better than any of the mxChoice
3735             ** paths currently in the best-so-far buffer.  So discard
3736             ** this candidate as not viable. */
3737 #ifdef WHERETRACE_ENABLED /* 0x4 */
3738             if( sqlite3WhereTrace&0x4 ){
3739               sqlite3DebugPrintf("Skip   %s cost=%-3d,%3d order=%c\n",
3740                   wherePathName(pFrom, iLoop, pWLoop), rCost, nOut,
3741                   isOrdered>=0 ? isOrdered+'0' : '?');
3742             }
3743 #endif
3744             continue;
3745           }
3746           /* If we reach this points it means that the new candidate path
3747           ** needs to be added to the set of best-so-far paths. */
3748           if( nTo<mxChoice ){
3749             /* Increase the size of the aTo set by one */
3750             jj = nTo++;
3751           }else{
3752             /* New path replaces the prior worst to keep count below mxChoice */
3753             jj = mxI;
3754           }
3755           pTo = &aTo[jj];
3756 #ifdef WHERETRACE_ENABLED /* 0x4 */
3757           if( sqlite3WhereTrace&0x4 ){
3758             sqlite3DebugPrintf("New    %s cost=%-3d,%3d order=%c\n",
3759                 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut,
3760                 isOrdered>=0 ? isOrdered+'0' : '?');
3761           }
3762 #endif
3763         }else{
3764           /* Control reaches here if best-so-far path pTo=aTo[jj] covers the
3765           ** same set of loops and has the sam isOrdered setting as the
3766           ** candidate path.  Check to see if the candidate should replace
3767           ** pTo or if the candidate should be skipped */
3768           if( pTo->rCost<rCost || (pTo->rCost==rCost && pTo->nRow<=nOut) ){
3769 #ifdef WHERETRACE_ENABLED /* 0x4 */
3770             if( sqlite3WhereTrace&0x4 ){
3771               sqlite3DebugPrintf(
3772                   "Skip   %s cost=%-3d,%3d order=%c",
3773                   wherePathName(pFrom, iLoop, pWLoop), rCost, nOut,
3774                   isOrdered>=0 ? isOrdered+'0' : '?');
3775               sqlite3DebugPrintf("   vs %s cost=%-3d,%d order=%c\n",
3776                   wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
3777                   pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?');
3778             }
3779 #endif
3780             /* Discard the candidate path from further consideration */
3781             testcase( pTo->rCost==rCost );
3782             continue;
3783           }
3784           testcase( pTo->rCost==rCost+1 );
3785           /* Control reaches here if the candidate path is better than the
3786           ** pTo path.  Replace pTo with the candidate. */
3787 #ifdef WHERETRACE_ENABLED /* 0x4 */
3788           if( sqlite3WhereTrace&0x4 ){
3789             sqlite3DebugPrintf(
3790                 "Update %s cost=%-3d,%3d order=%c",
3791                 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut,
3792                 isOrdered>=0 ? isOrdered+'0' : '?');
3793             sqlite3DebugPrintf("  was %s cost=%-3d,%3d order=%c\n",
3794                 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
3795                 pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?');
3796           }
3797 #endif
3798         }
3799         /* pWLoop is a winner.  Add it to the set of best so far */
3800         pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf;
3801         pTo->revLoop = revMask;
3802         pTo->nRow = nOut;
3803         pTo->rCost = rCost;
3804         pTo->rUnsorted = rUnsorted;
3805         pTo->isOrdered = isOrdered;
3806         memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop);
3807         pTo->aLoop[iLoop] = pWLoop;
3808         if( nTo>=mxChoice ){
3809           mxI = 0;
3810           mxCost = aTo[0].rCost;
3811           mxUnsorted = aTo[0].nRow;
3812           for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){
3813             if( pTo->rCost>mxCost
3814              || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted)
3815             ){
3816               mxCost = pTo->rCost;
3817               mxUnsorted = pTo->rUnsorted;
3818               mxI = jj;
3819             }
3820           }
3821         }
3822       }
3823     }
3824 
3825 #ifdef WHERETRACE_ENABLED  /* >=2 */
3826     if( sqlite3WhereTrace & 0x02 ){
3827       sqlite3DebugPrintf("---- after round %d ----\n", iLoop);
3828       for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){
3829         sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c",
3830            wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow,
3831            pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?');
3832         if( pTo->isOrdered>0 ){
3833           sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop);
3834         }else{
3835           sqlite3DebugPrintf("\n");
3836         }
3837       }
3838     }
3839 #endif
3840 
3841     /* Swap the roles of aFrom and aTo for the next generation */
3842     pFrom = aTo;
3843     aTo = aFrom;
3844     aFrom = pFrom;
3845     nFrom = nTo;
3846   }
3847 
3848   if( nFrom==0 ){
3849     sqlite3ErrorMsg(pParse, "no query solution");
3850     sqlite3DbFree(db, pSpace);
3851     return SQLITE_ERROR;
3852   }
3853 
3854   /* Find the lowest cost path.  pFrom will be left pointing to that path */
3855   pFrom = aFrom;
3856   for(ii=1; ii<nFrom; ii++){
3857     if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii];
3858   }
3859   assert( pWInfo->nLevel==nLoop );
3860   /* Load the lowest cost path into pWInfo */
3861   for(iLoop=0; iLoop<nLoop; iLoop++){
3862     WhereLevel *pLevel = pWInfo->a + iLoop;
3863     pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop];
3864     pLevel->iFrom = pWLoop->iTab;
3865     pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor;
3866   }
3867   if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0
3868    && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0
3869    && pWInfo->eDistinct==WHERE_DISTINCT_NOOP
3870    && nRowEst
3871   ){
3872     Bitmask notUsed;
3873     int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom,
3874                  WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], &notUsed);
3875     if( rc==pWInfo->pResultSet->nExpr ){
3876       pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
3877     }
3878   }
3879   if( pWInfo->pOrderBy ){
3880     if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){
3881       if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){
3882         pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
3883       }
3884     }else{
3885       pWInfo->nOBSat = pFrom->isOrdered;
3886       if( pWInfo->nOBSat<0 ) pWInfo->nOBSat = 0;
3887       pWInfo->revMask = pFrom->revLoop;
3888     }
3889     if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP)
3890         && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0
3891     ){
3892       Bitmask revMask = 0;
3893       int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy,
3894           pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask
3895       );
3896       assert( pWInfo->sorted==0 );
3897       if( nOrder==pWInfo->pOrderBy->nExpr ){
3898         pWInfo->sorted = 1;
3899         pWInfo->revMask = revMask;
3900       }
3901     }
3902   }
3903 
3904 
3905   pWInfo->nRowOut = pFrom->nRow;
3906 
3907   /* Free temporary memory and return success */
3908   sqlite3DbFree(db, pSpace);
3909   return SQLITE_OK;
3910 }
3911 
3912 /*
3913 ** Most queries use only a single table (they are not joins) and have
3914 ** simple == constraints against indexed fields.  This routine attempts
3915 ** to plan those simple cases using much less ceremony than the
3916 ** general-purpose query planner, and thereby yield faster sqlite3_prepare()
3917 ** times for the common case.
3918 **
3919 ** Return non-zero on success, if this query can be handled by this
3920 ** no-frills query planner.  Return zero if this query needs the
3921 ** general-purpose query planner.
3922 */
3923 static int whereShortCut(WhereLoopBuilder *pBuilder){
3924   WhereInfo *pWInfo;
3925   struct SrcList_item *pItem;
3926   WhereClause *pWC;
3927   WhereTerm *pTerm;
3928   WhereLoop *pLoop;
3929   int iCur;
3930   int j;
3931   Table *pTab;
3932   Index *pIdx;
3933 
3934   pWInfo = pBuilder->pWInfo;
3935   if( pWInfo->wctrlFlags & WHERE_FORCE_TABLE ) return 0;
3936   assert( pWInfo->pTabList->nSrc>=1 );
3937   pItem = pWInfo->pTabList->a;
3938   pTab = pItem->pTab;
3939   if( IsVirtual(pTab) ) return 0;
3940   if( pItem->fg.isIndexedBy ) return 0;
3941   iCur = pItem->iCursor;
3942   pWC = &pWInfo->sWC;
3943   pLoop = pBuilder->pNew;
3944   pLoop->wsFlags = 0;
3945   pLoop->nSkip = 0;
3946   pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0);
3947   if( pTerm ){
3948     testcase( pTerm->eOperator & WO_IS );
3949     pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW;
3950     pLoop->aLTerm[0] = pTerm;
3951     pLoop->nLTerm = 1;
3952     pLoop->u.btree.nEq = 1;
3953     /* TUNING: Cost of a rowid lookup is 10 */
3954     pLoop->rRun = 33;  /* 33==sqlite3LogEst(10) */
3955   }else{
3956     for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){
3957       int opMask;
3958       assert( pLoop->aLTermSpace==pLoop->aLTerm );
3959       if( !IsUniqueIndex(pIdx)
3960        || pIdx->pPartIdxWhere!=0
3961        || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace)
3962       ) continue;
3963       opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ;
3964       for(j=0; j<pIdx->nKeyCol; j++){
3965         pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx);
3966         if( pTerm==0 ) break;
3967         testcase( pTerm->eOperator & WO_IS );
3968         pLoop->aLTerm[j] = pTerm;
3969       }
3970       if( j!=pIdx->nKeyCol ) continue;
3971       pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED;
3972       if( pIdx->isCovering || (pItem->colUsed & ~columnsInIndex(pIdx))==0 ){
3973         pLoop->wsFlags |= WHERE_IDX_ONLY;
3974       }
3975       pLoop->nLTerm = j;
3976       pLoop->u.btree.nEq = j;
3977       pLoop->u.btree.pIndex = pIdx;
3978       /* TUNING: Cost of a unique index lookup is 15 */
3979       pLoop->rRun = 39;  /* 39==sqlite3LogEst(15) */
3980       break;
3981     }
3982   }
3983   if( pLoop->wsFlags ){
3984     pLoop->nOut = (LogEst)1;
3985     pWInfo->a[0].pWLoop = pLoop;
3986     pLoop->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur);
3987     pWInfo->a[0].iTabCur = iCur;
3988     pWInfo->nRowOut = 1;
3989     if( pWInfo->pOrderBy ) pWInfo->nOBSat =  pWInfo->pOrderBy->nExpr;
3990     if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){
3991       pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
3992     }
3993 #ifdef SQLITE_DEBUG
3994     pLoop->cId = '0';
3995 #endif
3996     return 1;
3997   }
3998   return 0;
3999 }
4000 
4001 /*
4002 ** Generate the beginning of the loop used for WHERE clause processing.
4003 ** The return value is a pointer to an opaque structure that contains
4004 ** information needed to terminate the loop.  Later, the calling routine
4005 ** should invoke sqlite3WhereEnd() with the return value of this function
4006 ** in order to complete the WHERE clause processing.
4007 **
4008 ** If an error occurs, this routine returns NULL.
4009 **
4010 ** The basic idea is to do a nested loop, one loop for each table in
4011 ** the FROM clause of a select.  (INSERT and UPDATE statements are the
4012 ** same as a SELECT with only a single table in the FROM clause.)  For
4013 ** example, if the SQL is this:
4014 **
4015 **       SELECT * FROM t1, t2, t3 WHERE ...;
4016 **
4017 ** Then the code generated is conceptually like the following:
4018 **
4019 **      foreach row1 in t1 do       \    Code generated
4020 **        foreach row2 in t2 do      |-- by sqlite3WhereBegin()
4021 **          foreach row3 in t3 do   /
4022 **            ...
4023 **          end                     \    Code generated
4024 **        end                        |-- by sqlite3WhereEnd()
4025 **      end                         /
4026 **
4027 ** Note that the loops might not be nested in the order in which they
4028 ** appear in the FROM clause if a different order is better able to make
4029 ** use of indices.  Note also that when the IN operator appears in
4030 ** the WHERE clause, it might result in additional nested loops for
4031 ** scanning through all values on the right-hand side of the IN.
4032 **
4033 ** There are Btree cursors associated with each table.  t1 uses cursor
4034 ** number pTabList->a[0].iCursor.  t2 uses the cursor pTabList->a[1].iCursor.
4035 ** And so forth.  This routine generates code to open those VDBE cursors
4036 ** and sqlite3WhereEnd() generates the code to close them.
4037 **
4038 ** The code that sqlite3WhereBegin() generates leaves the cursors named
4039 ** in pTabList pointing at their appropriate entries.  The [...] code
4040 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract
4041 ** data from the various tables of the loop.
4042 **
4043 ** If the WHERE clause is empty, the foreach loops must each scan their
4044 ** entire tables.  Thus a three-way join is an O(N^3) operation.  But if
4045 ** the tables have indices and there are terms in the WHERE clause that
4046 ** refer to those indices, a complete table scan can be avoided and the
4047 ** code will run much faster.  Most of the work of this routine is checking
4048 ** to see if there are indices that can be used to speed up the loop.
4049 **
4050 ** Terms of the WHERE clause are also used to limit which rows actually
4051 ** make it to the "..." in the middle of the loop.  After each "foreach",
4052 ** terms of the WHERE clause that use only terms in that loop and outer
4053 ** loops are evaluated and if false a jump is made around all subsequent
4054 ** inner loops (or around the "..." if the test occurs within the inner-
4055 ** most loop)
4056 **
4057 ** OUTER JOINS
4058 **
4059 ** An outer join of tables t1 and t2 is conceptally coded as follows:
4060 **
4061 **    foreach row1 in t1 do
4062 **      flag = 0
4063 **      foreach row2 in t2 do
4064 **        start:
4065 **          ...
4066 **          flag = 1
4067 **      end
4068 **      if flag==0 then
4069 **        move the row2 cursor to a null row
4070 **        goto start
4071 **      fi
4072 **    end
4073 **
4074 ** ORDER BY CLAUSE PROCESSING
4075 **
4076 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause
4077 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement
4078 ** if there is one.  If there is no ORDER BY clause or if this routine
4079 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL.
4080 **
4081 ** The iIdxCur parameter is the cursor number of an index.  If
4082 ** WHERE_ONETABLE_ONLY is set, iIdxCur is the cursor number of an index
4083 ** to use for OR clause processing.  The WHERE clause should use this
4084 ** specific cursor.  If WHERE_ONEPASS_DESIRED is set, then iIdxCur is
4085 ** the first cursor in an array of cursors for all indices.  iIdxCur should
4086 ** be used to compute the appropriate cursor depending on which index is
4087 ** used.
4088 */
4089 WhereInfo *sqlite3WhereBegin(
4090   Parse *pParse,        /* The parser context */
4091   SrcList *pTabList,    /* FROM clause: A list of all tables to be scanned */
4092   Expr *pWhere,         /* The WHERE clause */
4093   ExprList *pOrderBy,   /* An ORDER BY (or GROUP BY) clause, or NULL */
4094   ExprList *pResultSet, /* Result set of the query */
4095   u16 wctrlFlags,       /* One of the WHERE_* flags defined in sqliteInt.h */
4096   int iAuxArg           /* If WHERE_ONETABLE_ONLY is set, index cursor number,
4097                         ** If WHERE_USE_LIMIT, then the limit amount */
4098 ){
4099   int nByteWInfo;            /* Num. bytes allocated for WhereInfo struct */
4100   int nTabList;              /* Number of elements in pTabList */
4101   WhereInfo *pWInfo;         /* Will become the return value of this function */
4102   Vdbe *v = pParse->pVdbe;   /* The virtual database engine */
4103   Bitmask notReady;          /* Cursors that are not yet positioned */
4104   WhereLoopBuilder sWLB;     /* The WhereLoop builder */
4105   WhereMaskSet *pMaskSet;    /* The expression mask set */
4106   WhereLevel *pLevel;        /* A single level in pWInfo->a[] */
4107   WhereLoop *pLoop;          /* Pointer to a single WhereLoop object */
4108   int ii;                    /* Loop counter */
4109   sqlite3 *db;               /* Database connection */
4110   int rc;                    /* Return code */
4111   u8 bFordelete = 0;         /* OPFLAG_FORDELETE or zero, as appropriate */
4112 
4113   assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || (
4114         (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0
4115      && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0
4116   ));
4117 
4118   /* Only one of WHERE_ONETABLE_ONLY or WHERE_USE_LIMIT */
4119   assert( (wctrlFlags & WHERE_ONETABLE_ONLY)==0
4120             || (wctrlFlags & WHERE_USE_LIMIT)==0 );
4121 
4122   /* Variable initialization */
4123   db = pParse->db;
4124   memset(&sWLB, 0, sizeof(sWLB));
4125 
4126   /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */
4127   testcase( pOrderBy && pOrderBy->nExpr==BMS-1 );
4128   if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0;
4129   sWLB.pOrderBy = pOrderBy;
4130 
4131   /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via
4132   ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */
4133   if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){
4134     wctrlFlags &= ~WHERE_WANT_DISTINCT;
4135   }
4136 
4137   /* The number of tables in the FROM clause is limited by the number of
4138   ** bits in a Bitmask
4139   */
4140   testcase( pTabList->nSrc==BMS );
4141   if( pTabList->nSrc>BMS ){
4142     sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS);
4143     return 0;
4144   }
4145 
4146   /* This function normally generates a nested loop for all tables in
4147   ** pTabList.  But if the WHERE_ONETABLE_ONLY flag is set, then we should
4148   ** only generate code for the first table in pTabList and assume that
4149   ** any cursors associated with subsequent tables are uninitialized.
4150   */
4151   nTabList = (wctrlFlags & WHERE_ONETABLE_ONLY) ? 1 : pTabList->nSrc;
4152 
4153   /* Allocate and initialize the WhereInfo structure that will become the
4154   ** return value. A single allocation is used to store the WhereInfo
4155   ** struct, the contents of WhereInfo.a[], the WhereClause structure
4156   ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte
4157   ** field (type Bitmask) it must be aligned on an 8-byte boundary on
4158   ** some architectures. Hence the ROUND8() below.
4159   */
4160   nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel));
4161   pWInfo = sqlite3DbMallocZero(db, nByteWInfo + sizeof(WhereLoop));
4162   if( db->mallocFailed ){
4163     sqlite3DbFree(db, pWInfo);
4164     pWInfo = 0;
4165     goto whereBeginError;
4166   }
4167   pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1;
4168   pWInfo->nLevel = nTabList;
4169   pWInfo->pParse = pParse;
4170   pWInfo->pTabList = pTabList;
4171   pWInfo->pOrderBy = pOrderBy;
4172   pWInfo->pResultSet = pResultSet;
4173   pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(v);
4174   pWInfo->wctrlFlags = wctrlFlags;
4175   pWInfo->iLimit = iAuxArg;
4176   pWInfo->savedNQueryLoop = pParse->nQueryLoop;
4177   assert( pWInfo->eOnePass==ONEPASS_OFF );  /* ONEPASS defaults to OFF */
4178   pMaskSet = &pWInfo->sMaskSet;
4179   sWLB.pWInfo = pWInfo;
4180   sWLB.pWC = &pWInfo->sWC;
4181   sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo);
4182   assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) );
4183   whereLoopInit(sWLB.pNew);
4184 #ifdef SQLITE_DEBUG
4185   sWLB.pNew->cId = '*';
4186 #endif
4187 
4188   /* Split the WHERE clause into separate subexpressions where each
4189   ** subexpression is separated by an AND operator.
4190   */
4191   initMaskSet(pMaskSet);
4192   sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo);
4193   sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND);
4194 
4195   /* Special case: a WHERE clause that is constant.  Evaluate the
4196   ** expression and either jump over all of the code or fall thru.
4197   */
4198   for(ii=0; ii<sWLB.pWC->nTerm; ii++){
4199     if( nTabList==0 || sqlite3ExprIsConstantNotJoin(sWLB.pWC->a[ii].pExpr) ){
4200       sqlite3ExprIfFalse(pParse, sWLB.pWC->a[ii].pExpr, pWInfo->iBreak,
4201                          SQLITE_JUMPIFNULL);
4202       sWLB.pWC->a[ii].wtFlags |= TERM_CODED;
4203     }
4204   }
4205 
4206   /* Special case: No FROM clause
4207   */
4208   if( nTabList==0 ){
4209     if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr;
4210     if( wctrlFlags & WHERE_WANT_DISTINCT ){
4211       pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
4212     }
4213   }
4214 
4215   /* Assign a bit from the bitmask to every term in the FROM clause.
4216   **
4217   ** The N-th term of the FROM clause is assigned a bitmask of 1<<N.
4218   **
4219   ** The rule of the previous sentence ensures thta if X is the bitmask for
4220   ** a table T, then X-1 is the bitmask for all other tables to the left of T.
4221   ** Knowing the bitmask for all tables to the left of a left join is
4222   ** important.  Ticket #3015.
4223   **
4224   ** Note that bitmasks are created for all pTabList->nSrc tables in
4225   ** pTabList, not just the first nTabList tables.  nTabList is normally
4226   ** equal to pTabList->nSrc but might be shortened to 1 if the
4227   ** WHERE_ONETABLE_ONLY flag is set.
4228   */
4229   for(ii=0; ii<pTabList->nSrc; ii++){
4230     createMask(pMaskSet, pTabList->a[ii].iCursor);
4231     sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC);
4232   }
4233 #ifdef SQLITE_DEBUG
4234   for(ii=0; ii<pTabList->nSrc; ii++){
4235     Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor);
4236     assert( m==MASKBIT(ii) );
4237   }
4238 #endif
4239 
4240   /* Analyze all of the subexpressions. */
4241   sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC);
4242   if( db->mallocFailed ) goto whereBeginError;
4243 
4244   if( wctrlFlags & WHERE_WANT_DISTINCT ){
4245     if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){
4246       /* The DISTINCT marking is pointless.  Ignore it. */
4247       pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE;
4248     }else if( pOrderBy==0 ){
4249       /* Try to ORDER BY the result set to make distinct processing easier */
4250       pWInfo->wctrlFlags |= WHERE_DISTINCTBY;
4251       pWInfo->pOrderBy = pResultSet;
4252     }
4253   }
4254 
4255   /* Construct the WhereLoop objects */
4256 #if defined(WHERETRACE_ENABLED)
4257   if( sqlite3WhereTrace & 0xffff ){
4258     sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags);
4259     if( wctrlFlags & WHERE_USE_LIMIT ){
4260       sqlite3DebugPrintf(", limit: %d", iAuxArg);
4261     }
4262     sqlite3DebugPrintf(")\n");
4263   }
4264   if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */
4265     int i;
4266     for(i=0; i<sWLB.pWC->nTerm; i++){
4267       whereTermPrint(&sWLB.pWC->a[i], i);
4268     }
4269   }
4270 #endif
4271 
4272   if( nTabList!=1 || whereShortCut(&sWLB)==0 ){
4273     rc = whereLoopAddAll(&sWLB);
4274     if( rc ) goto whereBeginError;
4275 
4276 #ifdef WHERETRACE_ENABLED
4277     if( sqlite3WhereTrace ){    /* Display all of the WhereLoop objects */
4278       WhereLoop *p;
4279       int i;
4280       static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz"
4281                                              "ABCDEFGHIJKLMNOPQRSTUVWYXZ";
4282       for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){
4283         p->cId = zLabel[i%sizeof(zLabel)];
4284         whereLoopPrint(p, sWLB.pWC);
4285       }
4286     }
4287 #endif
4288 
4289     wherePathSolver(pWInfo, 0);
4290     if( db->mallocFailed ) goto whereBeginError;
4291     if( pWInfo->pOrderBy ){
4292        wherePathSolver(pWInfo, pWInfo->nRowOut+1);
4293        if( db->mallocFailed ) goto whereBeginError;
4294     }
4295   }
4296   if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){
4297      pWInfo->revMask = ALLBITS;
4298   }
4299   if( pParse->nErr || NEVER(db->mallocFailed) ){
4300     goto whereBeginError;
4301   }
4302 #ifdef WHERETRACE_ENABLED
4303   if( sqlite3WhereTrace ){
4304     sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut);
4305     if( pWInfo->nOBSat>0 ){
4306       sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask);
4307     }
4308     switch( pWInfo->eDistinct ){
4309       case WHERE_DISTINCT_UNIQUE: {
4310         sqlite3DebugPrintf("  DISTINCT=unique");
4311         break;
4312       }
4313       case WHERE_DISTINCT_ORDERED: {
4314         sqlite3DebugPrintf("  DISTINCT=ordered");
4315         break;
4316       }
4317       case WHERE_DISTINCT_UNORDERED: {
4318         sqlite3DebugPrintf("  DISTINCT=unordered");
4319         break;
4320       }
4321     }
4322     sqlite3DebugPrintf("\n");
4323     for(ii=0; ii<pWInfo->nLevel; ii++){
4324       whereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC);
4325     }
4326   }
4327 #endif
4328   /* Attempt to omit tables from the join that do not effect the result */
4329   if( pWInfo->nLevel>=2
4330    && pResultSet!=0
4331    && OptimizationEnabled(db, SQLITE_OmitNoopJoin)
4332   ){
4333     Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet);
4334     if( sWLB.pOrderBy ){
4335       tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy);
4336     }
4337     while( pWInfo->nLevel>=2 ){
4338       WhereTerm *pTerm, *pEnd;
4339       pLoop = pWInfo->a[pWInfo->nLevel-1].pWLoop;
4340       if( (pWInfo->pTabList->a[pLoop->iTab].fg.jointype & JT_LEFT)==0 ) break;
4341       if( (wctrlFlags & WHERE_WANT_DISTINCT)==0
4342        && (pLoop->wsFlags & WHERE_ONEROW)==0
4343       ){
4344         break;
4345       }
4346       if( (tabUsed & pLoop->maskSelf)!=0 ) break;
4347       pEnd = sWLB.pWC->a + sWLB.pWC->nTerm;
4348       for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){
4349         if( (pTerm->prereqAll & pLoop->maskSelf)!=0
4350          && !ExprHasProperty(pTerm->pExpr, EP_FromJoin)
4351         ){
4352           break;
4353         }
4354       }
4355       if( pTerm<pEnd ) break;
4356       WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId));
4357       pWInfo->nLevel--;
4358       nTabList--;
4359     }
4360   }
4361   WHERETRACE(0xffff,("*** Optimizer Finished ***\n"));
4362   pWInfo->pParse->nQueryLoop += pWInfo->nRowOut;
4363 
4364   /* If the caller is an UPDATE or DELETE statement that is requesting
4365   ** to use a one-pass algorithm, determine if this is appropriate.
4366   */
4367   assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 );
4368   if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){
4369     int wsFlags = pWInfo->a[0].pWLoop->wsFlags;
4370     int bOnerow = (wsFlags & WHERE_ONEROW)!=0;
4371     if( bOnerow
4372      || ((wctrlFlags & WHERE_ONEPASS_MULTIROW)!=0
4373            && 0==(wsFlags & WHERE_VIRTUALTABLE))
4374     ){
4375       pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI;
4376       if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){
4377         if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){
4378           bFordelete = OPFLAG_FORDELETE;
4379         }
4380         pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY);
4381       }
4382     }
4383   }
4384 
4385   /* Open all tables in the pTabList and any indices selected for
4386   ** searching those tables.
4387   */
4388   for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){
4389     Table *pTab;     /* Table to open */
4390     int iDb;         /* Index of database containing table/index */
4391     struct SrcList_item *pTabItem;
4392 
4393     pTabItem = &pTabList->a[pLevel->iFrom];
4394     pTab = pTabItem->pTab;
4395     iDb = sqlite3SchemaToIndex(db, pTab->pSchema);
4396     pLoop = pLevel->pWLoop;
4397     if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){
4398       /* Do nothing */
4399     }else
4400 #ifndef SQLITE_OMIT_VIRTUALTABLE
4401     if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){
4402       const char *pVTab = (const char *)sqlite3GetVTable(db, pTab);
4403       int iCur = pTabItem->iCursor;
4404       sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB);
4405     }else if( IsVirtual(pTab) ){
4406       /* noop */
4407     }else
4408 #endif
4409     if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0
4410          && (wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0 ){
4411       int op = OP_OpenRead;
4412       if( pWInfo->eOnePass!=ONEPASS_OFF ){
4413         op = OP_OpenWrite;
4414         pWInfo->aiCurOnePass[0] = pTabItem->iCursor;
4415       };
4416       sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op);
4417       assert( pTabItem->iCursor==pLevel->iTabCur );
4418       testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 );
4419       testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS );
4420       if( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol<BMS && HasRowid(pTab) ){
4421         Bitmask b = pTabItem->colUsed;
4422         int n = 0;
4423         for(; b; b=b>>1, n++){}
4424         sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32);
4425         assert( n<=pTab->nCol );
4426       }
4427 #ifdef SQLITE_ENABLE_CURSOR_HINTS
4428       if( pLoop->u.btree.pIndex!=0 ){
4429         sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete);
4430       }else
4431 #endif
4432       {
4433         sqlite3VdbeChangeP5(v, bFordelete);
4434       }
4435 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK
4436       sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0,
4437                             (const u8*)&pTabItem->colUsed, P4_INT64);
4438 #endif
4439     }else{
4440       sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName);
4441     }
4442     if( pLoop->wsFlags & WHERE_INDEXED ){
4443       Index *pIx = pLoop->u.btree.pIndex;
4444       int iIndexCur;
4445       int op = OP_OpenRead;
4446       /* iAuxArg is always set if to a positive value if ONEPASS is possible */
4447       assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 );
4448       if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx)
4449        && (wctrlFlags & WHERE_ONETABLE_ONLY)!=0
4450       ){
4451         /* This is one term of an OR-optimization using the PRIMARY KEY of a
4452         ** WITHOUT ROWID table.  No need for a separate index */
4453         iIndexCur = pLevel->iTabCur;
4454         op = 0;
4455       }else if( pWInfo->eOnePass!=ONEPASS_OFF ){
4456         Index *pJ = pTabItem->pTab->pIndex;
4457         iIndexCur = iAuxArg;
4458         assert( wctrlFlags & WHERE_ONEPASS_DESIRED );
4459         while( ALWAYS(pJ) && pJ!=pIx ){
4460           iIndexCur++;
4461           pJ = pJ->pNext;
4462         }
4463         op = OP_OpenWrite;
4464         pWInfo->aiCurOnePass[1] = iIndexCur;
4465       }else if( iAuxArg && (wctrlFlags & WHERE_ONETABLE_ONLY)!=0 ){
4466         iIndexCur = iAuxArg;
4467         if( wctrlFlags & WHERE_REOPEN_IDX ) op = OP_ReopenIdx;
4468       }else{
4469         iIndexCur = pParse->nTab++;
4470       }
4471       pLevel->iIdxCur = iIndexCur;
4472       assert( pIx->pSchema==pTab->pSchema );
4473       assert( iIndexCur>=0 );
4474       if( op ){
4475         sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb);
4476         sqlite3VdbeSetP4KeyInfo(pParse, pIx);
4477         if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0
4478          && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0
4479          && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0
4480         ){
4481           sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); /* Hint to COMDB2 */
4482         }
4483         VdbeComment((v, "%s", pIx->zName));
4484 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK
4485         {
4486           u64 colUsed = 0;
4487           int ii, jj;
4488           for(ii=0; ii<pIx->nColumn; ii++){
4489             jj = pIx->aiColumn[ii];
4490             if( jj<0 ) continue;
4491             if( jj>63 ) jj = 63;
4492             if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue;
4493             colUsed |= ((u64)1)<<(ii<63 ? ii : 63);
4494           }
4495           sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0,
4496                                 (u8*)&colUsed, P4_INT64);
4497         }
4498 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */
4499       }
4500     }
4501     if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb);
4502   }
4503   pWInfo->iTop = sqlite3VdbeCurrentAddr(v);
4504   if( db->mallocFailed ) goto whereBeginError;
4505 
4506   /* Generate the code to do the search.  Each iteration of the for
4507   ** loop below generates code for a single nested loop of the VM
4508   ** program.
4509   */
4510   notReady = ~(Bitmask)0;
4511   for(ii=0; ii<nTabList; ii++){
4512     int addrExplain;
4513     int wsFlags;
4514     pLevel = &pWInfo->a[ii];
4515     wsFlags = pLevel->pWLoop->wsFlags;
4516 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX
4517     if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){
4518       constructAutomaticIndex(pParse, &pWInfo->sWC,
4519                 &pTabList->a[pLevel->iFrom], notReady, pLevel);
4520       if( db->mallocFailed ) goto whereBeginError;
4521     }
4522 #endif
4523     addrExplain = sqlite3WhereExplainOneScan(
4524         pParse, pTabList, pLevel, ii, pLevel->iFrom, wctrlFlags
4525     );
4526     pLevel->addrBody = sqlite3VdbeCurrentAddr(v);
4527     notReady = sqlite3WhereCodeOneLoopStart(pWInfo, ii, notReady);
4528     pWInfo->iContinue = pLevel->addrCont;
4529     if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_ONETABLE_ONLY)==0 ){
4530       sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain);
4531     }
4532   }
4533 
4534   /* Done. */
4535   VdbeModuleComment((v, "Begin WHERE-core"));
4536   return pWInfo;
4537 
4538   /* Jump here if malloc fails */
4539 whereBeginError:
4540   if( pWInfo ){
4541     pParse->nQueryLoop = pWInfo->savedNQueryLoop;
4542     whereInfoFree(db, pWInfo);
4543   }
4544   return 0;
4545 }
4546 
4547 /*
4548 ** Generate the end of the WHERE loop.  See comments on
4549 ** sqlite3WhereBegin() for additional information.
4550 */
4551 void sqlite3WhereEnd(WhereInfo *pWInfo){
4552   Parse *pParse = pWInfo->pParse;
4553   Vdbe *v = pParse->pVdbe;
4554   int i;
4555   WhereLevel *pLevel;
4556   WhereLoop *pLoop;
4557   SrcList *pTabList = pWInfo->pTabList;
4558   sqlite3 *db = pParse->db;
4559 
4560   /* Generate loop termination code.
4561   */
4562   VdbeModuleComment((v, "End WHERE-core"));
4563   sqlite3ExprCacheClear(pParse);
4564   for(i=pWInfo->nLevel-1; i>=0; i--){
4565     int addr;
4566     pLevel = &pWInfo->a[i];
4567     pLoop = pLevel->pWLoop;
4568     sqlite3VdbeResolveLabel(v, pLevel->addrCont);
4569     if( pLevel->op!=OP_Noop ){
4570       sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3);
4571       sqlite3VdbeChangeP5(v, pLevel->p5);
4572       VdbeCoverage(v);
4573       VdbeCoverageIf(v, pLevel->op==OP_Next);
4574       VdbeCoverageIf(v, pLevel->op==OP_Prev);
4575       VdbeCoverageIf(v, pLevel->op==OP_VNext);
4576     }
4577     if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){
4578       struct InLoop *pIn;
4579       int j;
4580       sqlite3VdbeResolveLabel(v, pLevel->addrNxt);
4581       for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){
4582         sqlite3VdbeJumpHere(v, pIn->addrInTop+1);
4583         sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop);
4584         VdbeCoverage(v);
4585         VdbeCoverageIf(v, pIn->eEndLoopOp==OP_PrevIfOpen);
4586         VdbeCoverageIf(v, pIn->eEndLoopOp==OP_NextIfOpen);
4587         sqlite3VdbeJumpHere(v, pIn->addrInTop-1);
4588       }
4589     }
4590     sqlite3VdbeResolveLabel(v, pLevel->addrBrk);
4591     if( pLevel->addrSkip ){
4592       sqlite3VdbeGoto(v, pLevel->addrSkip);
4593       VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName));
4594       sqlite3VdbeJumpHere(v, pLevel->addrSkip);
4595       sqlite3VdbeJumpHere(v, pLevel->addrSkip-2);
4596     }
4597 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS
4598     if( pLevel->addrLikeRep ){
4599       int op;
4600       if( sqlite3VdbeGetOp(v, pLevel->addrLikeRep-1)->p1 ){
4601         op = OP_DecrJumpZero;
4602       }else{
4603         op = OP_JumpZeroIncr;
4604       }
4605       sqlite3VdbeAddOp2(v, op, pLevel->iLikeRepCntr, pLevel->addrLikeRep);
4606       VdbeCoverage(v);
4607     }
4608 #endif
4609     if( pLevel->iLeftJoin ){
4610       addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v);
4611       assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0
4612            || (pLoop->wsFlags & WHERE_INDEXED)!=0 );
4613       if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 ){
4614         sqlite3VdbeAddOp1(v, OP_NullRow, pTabList->a[i].iCursor);
4615       }
4616       if( pLoop->wsFlags & WHERE_INDEXED ){
4617         sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur);
4618       }
4619       if( pLevel->op==OP_Return ){
4620         sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst);
4621       }else{
4622         sqlite3VdbeGoto(v, pLevel->addrFirst);
4623       }
4624       sqlite3VdbeJumpHere(v, addr);
4625     }
4626     VdbeModuleComment((v, "End WHERE-loop%d: %s", i,
4627                      pWInfo->pTabList->a[pLevel->iFrom].pTab->zName));
4628   }
4629 
4630   /* The "break" point is here, just past the end of the outer loop.
4631   ** Set it.
4632   */
4633   sqlite3VdbeResolveLabel(v, pWInfo->iBreak);
4634 
4635   assert( pWInfo->nLevel<=pTabList->nSrc );
4636   for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){
4637     int k, last;
4638     VdbeOp *pOp;
4639     Index *pIdx = 0;
4640     struct SrcList_item *pTabItem = &pTabList->a[pLevel->iFrom];
4641     Table *pTab = pTabItem->pTab;
4642     assert( pTab!=0 );
4643     pLoop = pLevel->pWLoop;
4644 
4645     /* For a co-routine, change all OP_Column references to the table of
4646     ** the co-routine into OP_Copy of result contained in a register.
4647     ** OP_Rowid becomes OP_Null.
4648     */
4649     if( pTabItem->fg.viaCoroutine && !db->mallocFailed ){
4650       translateColumnToCopy(v, pLevel->addrBody, pLevel->iTabCur,
4651                             pTabItem->regResult, 0);
4652       continue;
4653     }
4654 
4655     /* Close all of the cursors that were opened by sqlite3WhereBegin.
4656     ** Except, do not close cursors that will be reused by the OR optimization
4657     ** (WHERE_OMIT_OPEN_CLOSE).  And do not close the OP_OpenWrite cursors
4658     ** created for the ONEPASS optimization.
4659     */
4660     if( (pTab->tabFlags & TF_Ephemeral)==0
4661      && pTab->pSelect==0
4662      && (pWInfo->wctrlFlags & WHERE_OMIT_OPEN_CLOSE)==0
4663     ){
4664       int ws = pLoop->wsFlags;
4665       if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){
4666         sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor);
4667       }
4668       if( (ws & WHERE_INDEXED)!=0
4669        && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0
4670        && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1]
4671       ){
4672         sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur);
4673       }
4674     }
4675 
4676     /* If this scan uses an index, make VDBE code substitutions to read data
4677     ** from the index instead of from the table where possible.  In some cases
4678     ** this optimization prevents the table from ever being read, which can
4679     ** yield a significant performance boost.
4680     **
4681     ** Calls to the code generator in between sqlite3WhereBegin and
4682     ** sqlite3WhereEnd will have created code that references the table
4683     ** directly.  This loop scans all that code looking for opcodes
4684     ** that reference the table and converts them into opcodes that
4685     ** reference the index.
4686     */
4687     if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){
4688       pIdx = pLoop->u.btree.pIndex;
4689     }else if( pLoop->wsFlags & WHERE_MULTI_OR ){
4690       pIdx = pLevel->u.pCovidx;
4691     }
4692     if( pIdx
4693      && (pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable))
4694      && !db->mallocFailed
4695     ){
4696       last = sqlite3VdbeCurrentAddr(v);
4697       k = pLevel->addrBody;
4698       pOp = sqlite3VdbeGetOp(v, k);
4699       for(; k<last; k++, pOp++){
4700         if( pOp->p1!=pLevel->iTabCur ) continue;
4701         if( pOp->opcode==OP_Column ){
4702           int x = pOp->p2;
4703           assert( pIdx->pTable==pTab );
4704           if( !HasRowid(pTab) ){
4705             Index *pPk = sqlite3PrimaryKeyIndex(pTab);
4706             x = pPk->aiColumn[x];
4707             assert( x>=0 );
4708           }
4709           x = sqlite3ColumnOfIndex(pIdx, x);
4710           if( x>=0 ){
4711             pOp->p2 = x;
4712             pOp->p1 = pLevel->iIdxCur;
4713           }
4714           assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 );
4715         }else if( pOp->opcode==OP_Rowid ){
4716           pOp->p1 = pLevel->iIdxCur;
4717           pOp->opcode = OP_IdxRowid;
4718         }
4719       }
4720     }
4721   }
4722 
4723   /* Final cleanup
4724   */
4725   pParse->nQueryLoop = pWInfo->savedNQueryLoop;
4726   whereInfoFree(db, pWInfo);
4727   return;
4728 }
4729