SQLite

#define WHERE_INDEXED      0x000f0000  /* Anything that uses an index */
#define WHERE_NOT_FULLSCAN 0x100f3000  /* Does not do a full table scan */
#define WHERE_IN_ABLE      0x000f1000  /* Able to support an IN operator */
#define WHERE_TOP_LIMIT    0x00100000  /* x<EXPR or x<=EXPR constraint */
#define WHERE_BTM_LIMIT    0x00200000  /* x>EXPR or x>=EXPR constraint */
#define WHERE_BOTH_LIMIT   0x00300000  /* Both x>EXPR and x<EXPR */
#define WHERE_IDX_ONLY     0x00400000  /* Use index only - omit table */
#define WHERE_ORDERED      0x00800000  /* Output will appear in correct order */
#define WHERE_REVERSE      0x01000000  /* Scan in reverse order */
#define WHERE_UNIQUE       0x02000000  /* Selects no more than one row */
#define WHERE_ALL_UNIQUE   0x04000000  /* This and all prior have one row */
#define WHERE_VIRTUALTABLE 0x08000000  /* Use virtual-table processing */
#define WHERE_MULTI_OR     0x10000000  /* OR using multiple indices */
#define WHERE_TEMP_INDEX   0x20000000  /* Uses an ephemeral index */
#define WHERE_DISTINCT     0x40000000  /* Correct order for DISTINCT */

  ExprList *pOrderBy;             /* The ORDER BY clause */
  ExprList *pDistinct;            /* The select-list if query is DISTINCT */
  sqlite3_index_info **ppIdxInfo; /* Index information passed to xBestIndex */
  int i, n;                       /* Which loop is being coded; # of loops */
  WhereLevel *aLevel;             /* Info about outer loops */
  WhereCost cost;                 /* Lowest cost query plan */
};

/*
** Return TRUE if the probe cost is less than the baseline cost
*/
static int compareCost(const WhereCost *pProbe, const WhereCost *pBaseline){
  if( pProbe->rCost<pBaseline->rCost ) return 1;
  if( pProbe->rCost>pBaseline->rCost ) return 0;
  if( pProbe->plan.nOBSat>pBaseline->plan.nOBSat ) return 1;
  if( pProbe->plan.nRow<pBaseline->plan.nRow ) return 1;
  return 0;
}

/*
** Initialize a preallocated WhereClause structure.
*/
static void whereClauseInit(
  WhereClause *pWC,        /* The WhereClause to be initialized */
  Parse *pParse,           /* The parsing context */

      ** less than the current cost stored in pCost, replace the contents
      ** of pCost. */
      WHERETRACE(("... multi-index OR cost=%.9g nrow=%.9g\n", rTotal, nRow));
      if( rTotal<p->cost.rCost ){
        p->cost.rCost = rTotal;
        p->cost.used = used;
        p->cost.plan.nRow = nRow;
        p->cost.plan.nOBSat = p->i ? p->aLevel[p->i-1].plan.nOBSat : 0;
        p->cost.plan.wsFlags = flags;
        p->cost.plan.u.pTerm = pTerm;
      }
    }
  }
#endif /* SQLITE_OMIT_OR_OPTIMIZATION */
}

  if( (SQLITE_BIG_DBL/((double)2))<rCost ){
    p->cost.rCost = (SQLITE_BIG_DBL/((double)2));
  }else{
    p->cost.rCost = rCost;
  }
  p->cost.plan.u.pVtabIdx = pIdxInfo;
  if( pIdxInfo->orderByConsumed ){
    p->cost.plan.wsFlags |= WHERE_ORDERED;
    p->cost.plan.nOBSat = nOrderBy;
  }else{
    p->cost.plan.nOBSat = p->i ? p->aLevel[p->i-1].plan.nOBSat : 0;
  }
  p->cost.plan.nEq = 0;
  pIdxInfo->nOrderBy = nOrderBy;

  /* Try to find a more efficient access pattern by using multiple indexes
  ** to optimize an OR expression within the WHERE clause. 
  */

  Index *pIdx;
  u8 sortOrder;
  for(i=p->i-1; i>=0; i--, pLevel--){
    if( pLevel->iTabCur!=iTab ) continue;
    if( (pLevel->plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ){
      return 1;
    }
    if( (pLevel->plan.wsFlags & WHERE_ORDERED)==0 ){
      return 0;
    }
    if( (pIdx = pLevel->plan.u.pIdx)!=0 ){
      if( iCol<0 ){
        sortOrder = 0;
        testcase( (pLevel->plan.wsFlags & WHERE_REVERSE)!=0 );
      }else{
        int n = pIdx->nColumn;
        for(j=0; j<n; j++){
          if( iCol==pIdx->aiColumn[j] ) break;

  if( p->i==0 ){
    nPriorSat = 0;
  }else{
    nPriorSat = p->aLevel[p->i-1].plan.nOBSat;
    if( OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return nPriorSat;
  }
  if( nEqCol==0 ){
    if( p->i && (p->aLevel[p->i-1].plan.wsFlags & WHERE_ORDERED)==0 ){
      return nPriorSat;
    }
    nEqOneRow = 0;
  }else if( p->i==0 || (p->aLevel[p->i-1].plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ){
    nEqOneRow = nEqCol;
  }else{

    sortOrder = bOuterRev;
    nEqOneRow = -1;
  }
  pOrderBy = p->pOrderBy;
  assert( pOrderBy!=0 );
  if( wsFlags & WHERE_COLUMN_IN ) return nPriorSat;
  if( pIdx->bUnordered ) return nPriorSat;

    pIdx = 0;
  }

  /* Loop over all indices looking for the best one to use
  */
  for(; pProbe; pIdx=pProbe=pProbe->pNext){
    const tRowcnt * const aiRowEst = pProbe->aiRowEst;
    WhereCost pc;               /* Cost of using pProbe */

    double log10N = (double)1;  /* base-10 logarithm of nRow (inexact) */
    int bRev = 2;               /* 0=forward scan.  1=reverse.  2=undecided */
    memset(&pc, 0, sizeof(pc));


    /* The following variables are populated based on the properties of
    ** index being evaluated. They are then used to determine the expected
    ** cost and number of rows returned.
    **
    **  pc.plan.nEq: 
    **    Number of equality terms that can be implemented using the index.
    **    In other words, the number of initial fields in the index that
    **    are used in == or IN or NOT NULL constraints of the WHERE clause.
    **
    **  nInMul:  
    **    The "in-multiplier". This is an estimate of how many seek operations 
    **    SQLite must perform on the index in question. For example, if the 

    **    two queries requires table b-tree lookups in order to find the value
    **    of column c, but the first does not because columns a and b are
    **    both available in the index.
    **
    **             SELECT a, b    FROM tbl WHERE a = 1;
    **             SELECT a, b, c FROM tbl WHERE a = 1;
    */

    int nOrdered;                 /* Number of ordered terms matching index */
    int bInEst = 0;               /* True if "x IN (SELECT...)" seen */
    int nInMul = 1;               /* Number of distinct equalities to lookup */
    double rangeDiv = (double)1;  /* Estimated reduction in search space */
    int nBound = 0;               /* Number of range constraints seen */
    int bSort;                    /* True if external sort required */
    int bDist;                    /* True if index cannot help with DISTINCT */
    int bLookup = 0;              /* True if not a covering index */
    int nPriorSat;                /* ORDER BY terms satisfied by outer loops */
    int nOrderBy;                 /* Number of ORDER BY terms */
    WhereTerm *pTerm;             /* A single term of the WHERE clause */
#ifdef SQLITE_ENABLE_STAT3
    WhereTerm *pFirstTerm = 0;    /* First term matching the index */
#endif

    nOrderBy = p->pOrderBy ? p->pOrderBy->nExpr : 0;
    if( p->i ){
      nPriorSat = pc.plan.nOBSat = p->aLevel[p->i-1].plan.nOBSat;
      bSort = nPriorSat<nOrderBy;
      bDist = 0;
    }else{
      nPriorSat = pc.plan.nOBSat = 0;
      bSort = nOrderBy>0;
      bDist = p->pDistinct!=0;
    }

    /* Determine the values of pc.plan.nEq and nInMul */
    for(pc.plan.nEq=nOrdered=0; pc.plan.nEq<pProbe->nColumn; pc.plan.nEq++){
      int j = pProbe->aiColumn[pc.plan.nEq];
      pTerm = findTerm(pWC, iCur, j, p->notReady, eqTermMask, pIdx);
      if( pTerm==0 ) break;
      pc.plan.wsFlags |= (WHERE_COLUMN_EQ|WHERE_ROWID_EQ);
      testcase( pTerm->pWC!=pWC );
      if( pTerm->eOperator & WO_IN ){
        Expr *pExpr = pTerm->pExpr;
        pc.plan.wsFlags |= WHERE_COLUMN_IN;
        if( ExprHasProperty(pExpr, EP_xIsSelect) ){
          /* "x IN (SELECT ...)":  Assume the SELECT returns 25 rows */
          nInMul *= 25;
          bInEst = 1;
        }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){
          /* "x IN (value, value, ...)" */
          nInMul *= pExpr->x.pList->nExpr;
        }
      }else if( pTerm->eOperator & WO_ISNULL ){
        pc.plan.wsFlags |= WHERE_COLUMN_NULL;
        if( pc.plan.nEq==nOrdered ) nOrdered++;
      }else if( bSort && pc.plan.nEq==nOrdered && isOrderedTerm(p, pTerm, &bRev) ){
        nOrdered++;
      }
#ifdef SQLITE_ENABLE_STAT3
      if( pc.plan.nEq==0 && pProbe->aSample ) pFirstTerm = pTerm;
#endif
      pc.used |= pTerm->prereqRight;
    }
 
    /* If the index being considered is UNIQUE, and there is an equality 
    ** constraint for all columns in the index, then this search will find
    ** at most a single row. In this case set the WHERE_UNIQUE flag to 
    ** indicate this to the caller.
    **
    ** Otherwise, if the search may find more than one row, test to see if
    ** there is a range constraint on indexed column (pc.plan.nEq+1) that can be 
    ** optimized using the index. 
    */
    if( pc.plan.nEq==pProbe->nColumn && pProbe->onError!=OE_None ){
      testcase( pc.plan.wsFlags & WHERE_COLUMN_IN );
      testcase( pc.plan.wsFlags & WHERE_COLUMN_NULL );
      if( (pc.plan.wsFlags & (WHERE_COLUMN_IN|WHERE_COLUMN_NULL))==0 ){
        pc.plan.wsFlags |= WHERE_UNIQUE;
        if( p->i==0 || (p->aLevel[p->i-1].plan.wsFlags & WHERE_ALL_UNIQUE)!=0 ){
          pc.plan.wsFlags |= WHERE_ALL_UNIQUE;
        }
      }
    }else if( pProbe->bUnordered==0 ){
      int j;
      j = (pc.plan.nEq==pProbe->nColumn ? -1 : pProbe->aiColumn[pc.plan.nEq]);
      if( findTerm(pWC, iCur, j, p->notReady, WO_LT|WO_LE|WO_GT|WO_GE, pIdx) ){
        WhereTerm *pTop, *pBtm;
        pTop = findTerm(pWC, iCur, j, p->notReady, WO_LT|WO_LE, pIdx);
        pBtm = findTerm(pWC, iCur, j, p->notReady, WO_GT|WO_GE, pIdx);
        whereRangeScanEst(pParse, pProbe, pc.plan.nEq, pBtm, pTop, &rangeDiv);
        if( pTop ){
          nBound = 1;
          pc.plan.wsFlags |= WHERE_TOP_LIMIT;
          pc.used |= pTop->prereqRight;
          testcase( pTop->pWC!=pWC );
        }
        if( pBtm ){
          nBound++;
          pc.plan.wsFlags |= WHERE_BTM_LIMIT;
          pc.used |= pBtm->prereqRight;
          testcase( pBtm->pWC!=pWC );
        }
        pc.plan.wsFlags |= (WHERE_COLUMN_RANGE|WHERE_ROWID_RANGE);
      }
    }

    /* If there is an ORDER BY clause and the index being considered will
    ** naturally scan rows in the required order, set the appropriate flags
    ** in pc.plan.wsFlags. Otherwise, if there is an ORDER BY clause but
    ** the index will scan rows in a different order, set the bSort
    ** variable.  */
    assert( bRev>=0 && bRev<=2 );
    if( bSort ){
      testcase( bRev==0 );
      testcase( bRev==1 );
      testcase( bRev==2 );
      pc.plan.nOBSat = isSortingIndex(p, pProbe, iCur, nOrdered,
                                 pc.plan.wsFlags, bRev&1, &bRev);
      if( nPriorSat<pc.plan.nOBSat || (pc.plan.wsFlags & WHERE_UNIQUE)!=0 ){
        pc.plan.wsFlags |= WHERE_ORDERED;
      }
      if( nOrderBy==pc.plan.nOBSat ){
        bSort = 0;
        pc.plan.wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE;
      }
      if( bRev & 1 ) pc.plan.wsFlags |= WHERE_REVERSE;
    }

    /* If there is a DISTINCT qualifier and this index will scan rows in
    ** order of the DISTINCT expressions, clear bDist and set the appropriate
    ** flags in pc.plan.wsFlags. */
    if( bDist
     && isDistinctIndex(pParse, pWC, pProbe, iCur, p->pDistinct, pc.plan.nEq)
     && (pc.plan.wsFlags & WHERE_COLUMN_IN)==0
    ){
      bDist = 0;
      pc.plan.wsFlags |= WHERE_ROWID_RANGE|WHERE_COLUMN_RANGE|WHERE_DISTINCT;
    }

    /* If currently calculating the cost of using an index (not the IPK
    ** index), determine if all required column data may be obtained without 
    ** using the main table (i.e. if the index is a covering
    ** index for this query). If it is, set the WHERE_IDX_ONLY flag in
    ** pc.plan.wsFlags. Otherwise, set the bLookup variable to true.  */
    if( pIdx ){
      Bitmask m = pSrc->colUsed;
      int j;
      for(j=0; j<pIdx->nColumn; j++){
        int x = pIdx->aiColumn[j];
        if( x<BMS-1 ){
          m &= ~(((Bitmask)1)<<x);
        }
      }
      if( m==0 ){
        pc.plan.wsFlags |= WHERE_IDX_ONLY;
      }else{
        bLookup = 1;
      }
    }

    /*
    ** Estimate the number of rows of output.  For an "x IN (SELECT...)"
    ** constraint, do not let the estimate exceed half the rows in the table.
    */
    pc.plan.nRow = (double)(aiRowEst[pc.plan.nEq] * nInMul);
    if( bInEst && pc.plan.nRow*2>aiRowEst[0] ){
      pc.plan.nRow = aiRowEst[0]/2;
      nInMul = (int)(pc.plan.nRow / aiRowEst[pc.plan.nEq]);
    }

#ifdef SQLITE_ENABLE_STAT3
    /* If the constraint is of the form x=VALUE or x IN (E1,E2,...)
    ** and we do not think that values of x are unique and if histogram
    ** data is available for column x, then it might be possible
    ** to get a better estimate on the number of rows based on
    ** VALUE and how common that value is according to the histogram.
    */
    if( pc.plan.nRow>(double)1 && pc.plan.nEq==1
     && pFirstTerm!=0 && aiRowEst[1]>1 ){
      assert( (pFirstTerm->eOperator & (WO_EQ|WO_ISNULL|WO_IN))!=0 );
      if( pFirstTerm->eOperator & (WO_EQ|WO_ISNULL) ){
        testcase( pFirstTerm->eOperator==WO_EQ );
        testcase( pFirstTerm->eOperator==WO_ISNULL );
        whereEqualScanEst(pParse, pProbe, pFirstTerm->pExpr->pRight,
                          &pc.plan.nRow);
      }else if( bInEst==0 ){
        assert( pFirstTerm->eOperator==WO_IN );
        whereInScanEst(pParse, pProbe, pFirstTerm->pExpr->x.pList,
                       &pc.plan.nRow);
      }
    }
#endif /* SQLITE_ENABLE_STAT3 */

    /* Adjust the number of output rows and downward to reflect rows
    ** that are excluded by range constraints.
    */
    pc.plan.nRow = pc.plan.nRow/rangeDiv;
    if( pc.plan.nRow<1 ) pc.plan.nRow = 1;

    /* Experiments run on real SQLite databases show that the time needed
    ** to do a binary search to locate a row in a table or index is roughly
    ** log10(N) times the time to move from one row to the next row within
    ** a table or index.  The actual times can vary, with the size of
    ** records being an important factor.  Both moves and searches are
    ** slower with larger records, presumably because fewer records fit
    ** on one page and hence more pages have to be fetched.
    **
    ** The ANALYZE command and the sqlite_stat1 and sqlite_stat3 tables do
    ** not give us data on the relative sizes of table and index records.
    ** So this computation assumes table records are about twice as big
    ** as index records
    */
    if( (pc.plan.wsFlags&~(WHERE_REVERSE|WHERE_ORDERED))==WHERE_IDX_ONLY
     && (pWC->wctrlFlags & WHERE_ONEPASS_DESIRED)==0
     && sqlite3GlobalConfig.bUseCis
     && OptimizationEnabled(pParse->db, SQLITE_CoverIdxScan)
    ){
      /* This index is not useful for indexing, but it is a covering index.
      ** A full-scan of the index might be a little faster than a full-scan
      ** of the table, so give this case a cost slightly less than a table
      ** scan. */
      pc.rCost = aiRowEst[0]*3 + pProbe->nColumn;
      pc.plan.wsFlags |= WHERE_COVER_SCAN|WHERE_COLUMN_RANGE;
    }else if( (pc.plan.wsFlags & WHERE_NOT_FULLSCAN)==0 ){
      /* The cost of a full table scan is a number of move operations equal
      ** to the number of rows in the table.
      **
      ** We add an additional 4x penalty to full table scans.  This causes
      ** the cost function to err on the side of choosing an index over
      ** choosing a full scan.  This 4x full-scan penalty is an arguable
      ** decision and one which we expect to revisit in the future.  But
      ** it seems to be working well enough at the moment.
      */
      pc.rCost = aiRowEst[0]*4;
      pc.plan.wsFlags &= ~WHERE_IDX_ONLY;
    }else{
      log10N = estLog(aiRowEst[0]);
      pc.rCost = pc.plan.nRow;
      if( pIdx ){
        if( bLookup ){
          /* For an index lookup followed by a table lookup:
          **    nInMul index searches to find the start of each index range
          **  + nRow steps through the index
          **  + nRow table searches to lookup the table entry using the rowid
          */
          pc.rCost += (nInMul + pc.plan.nRow)*log10N;
        }else{
          /* For a covering index:
          **     nInMul index searches to find the initial entry 
          **   + nRow steps through the index
          */
          pc.rCost += nInMul*log10N;
        }
      }else{
        /* For a rowid primary key lookup:
        **    nInMult table searches to find the initial entry for each range
        **  + nRow steps through the table
        */
        pc.rCost += nInMul*log10N;
      }
    }

    /* Add in the estimated cost of sorting the result.  Actual experimental
    ** measurements of sorting performance in SQLite show that sorting time
    ** adds C*N*log10(N) to the cost, where N is the number of rows to be 
    ** sorted and C is a factor between 1.95 and 4.3.  We will split the
    ** difference and select C of 3.0.
    */
    if( bSort ){
      double m = estLog(pc.plan.nRow*(nOrderBy - pc.plan.nOBSat)/nOrderBy);
      m *= (double)(pc.plan.nOBSat ? 2 : 3);
      pc.rCost += pc.plan.nRow*m;
    }
    if( bDist ){
      pc.rCost += pc.plan.nRow*estLog(pc.plan.nRow)*3;
    }

    /**** Cost of using this index has now been computed ****/

    /* If there are additional constraints on this table that cannot
    ** be used with the current index, but which might lower the number
    ** of output rows, adjust the nRow value accordingly.  This only 
    ** matters if the current index is the least costly, so do not bother
    ** with this step if we already know this index will not be chosen.
    ** Also, never reduce the output row count below 2 using this step.
    **
    ** It is critical that the notValid mask be used here instead of
    ** the notReady mask.  When computing an "optimal" index, the notReady
    ** mask will only have one bit set - the bit for the current table.
    ** The notValid mask, on the other hand, always has all bits set for
    ** tables that are not in outer loops.  If notReady is used here instead
    ** of notValid, then a optimal index that depends on inner joins loops
    ** might be selected even when there exists an optimal index that has
    ** no such dependency.
    */
    if( pc.plan.nRow>2 && pc.rCost<=p->cost.rCost ){
      int k;                       /* Loop counter */
      int nSkipEq = pc.plan.nEq;   /* Number of == constraints to skip */
      int nSkipRange = nBound;     /* Number of < constraints to skip */
      Bitmask thisTab;             /* Bitmap for pSrc */

      thisTab = getMask(pWC->pMaskSet, iCur);
      for(pTerm=pWC->a, k=pWC->nTerm; pc.plan.nRow>2 && k; k--, pTerm++){
        if( pTerm->wtFlags & TERM_VIRTUAL ) continue;
        if( (pTerm->prereqAll & p->notValid)!=thisTab ) continue;
        if( pTerm->eOperator & (WO_EQ|WO_IN|WO_ISNULL) ){
          if( nSkipEq ){
            /* Ignore the first pc.plan.nEq equality matches since the index
            ** has already accounted for these */
            nSkipEq--;
          }else{
            /* Assume each additional equality match reduces the result
            ** set size by a factor of 10 */
            pc.plan.nRow /= 10;
          }
        }else if( pTerm->eOperator & (WO_LT|WO_LE|WO_GT|WO_GE) ){
          if( nSkipRange ){
            /* Ignore the first nSkipRange range constraints since the index
            ** has already accounted for these */
            nSkipRange--;
          }else{
            /* Assume each additional range constraint reduces the result
            ** set size by a factor of 3.  Indexed range constraints reduce
            ** the search space by a larger factor: 4.  We make indexed range
            ** more selective intentionally because of the subjective 
            ** observation that indexed range constraints really are more
            ** selective in practice, on average. */
            pc.plan.nRow /= 3;
          }
        }else if( pTerm->eOperator!=WO_NOOP ){
          /* Any other expression lowers the output row count by half */
          pc.plan.nRow /= 2;
        }
      }
      if( pc.plan.nRow<2 ) pc.plan.nRow = 2;
    }


    WHERETRACE((
      "%s(%s):\n"
      "    nEq=%d nInMul=%d rangeDiv=%d bSort=%d bLookup=%d wsFlags=0x%08x\n"
      "    notReady=0x%llx log10N=%.1f nRow=%.1f cost=%.1f\n"
      "    used=0x%llx nOrdered=%d nOBSat=%d\n",
      pSrc->pTab->zName, (pIdx ? pIdx->zName : "ipk"), 
      pc.plan.nEq, nInMul, (int)rangeDiv, bSort, bLookup, pc.plan.wsFlags,
      p->notReady, log10N, pc.plan.nRow, pc.rCost, pc.used, nOrdered,
      pc.plan.nOBSat
    ));

    /* If this index is the best we have seen so far, then record this
    ** index and its cost in the p->cost structure.
    */
    if( (!pIdx || pc.plan.wsFlags) && compareCost(&pc, &p->cost) ){



      p->cost = pc;

      p->cost.plan.wsFlags &= wsFlagMask;


      p->cost.plan.u.pIdx = pIdx;
    }

    /* If there was an INDEXED BY clause, then only that one index is
    ** considered. */
    if( pSrc->pIndex ) break;

  ** in. This is used for application testing, to help find cases
  ** where application behaviour depends on the (undefined) order that
  ** SQLite outputs rows in in the absence of an ORDER BY clause.  */
  if( !p->pOrderBy && pParse->db->flags & SQLITE_ReverseOrder ){
    p->cost.plan.wsFlags |= WHERE_REVERSE;
  }

  assert( p->pOrderBy || (p->cost.plan.wsFlags&WHERE_ORDERED)==0 );
  assert( p->cost.plan.u.pIdx==0 || (p->cost.plan.wsFlags&WHERE_ROWID_EQ)==0 );
  assert( pSrc->pIndex==0 
       || p->cost.plan.u.pIdx==0 
       || p->cost.plan.u.pIdx==pSrc->pIndex 
  );

  WHERETRACE(("best index is: %s\n",

         p->cost.plan.u.pIdx ? p->cost.plan.u.pIdx->zName : "ipk"));

  
  bestOrClauseIndex(p);
  bestAutomaticIndex(p);
  p->cost.plan.wsFlags |= eqTermMask;
}

/*

    ** query, then the caller will only allow the loop to run for
    ** a single iteration. This means that the first row returned
    ** should not have a NULL value stored in 'x'. If column 'x' is
    ** the first one after the nEq equality constraints in the index,
    ** this requires some special handling.
    */
    if( (wctrlFlags&WHERE_ORDERBY_MIN)!=0
     && (pLevel->plan.wsFlags&WHERE_ORDERED)
     && (pIdx->nColumn>nEq)
    ){
      /* assert( pOrderBy->nExpr==1 ); */
      /* assert( pOrderBy->a[0].pExpr->iColumn==pIdx->aiColumn[nEq] ); */
      isMinQuery = 1;
      nExtraReg = 1;
    }

        **       index specified by its INDEXED BY clause.  This rule ensures
        **       that a best-so-far is always selected even if an impossible
        **       combination of INDEXED BY clauses are given.  The error
        **       will be detected and relayed back to the application later.
        **       The NEVER() comes about because rule (2) above prevents
        **       An indexable full-table-scan from reaching rule (3).
        **
        **   (4) The plan cost must be lower than prior plans, where "cost"
        **       is defined by the compareCost() function above. 
        */
        if( (sWBI.cost.used&sWBI.notValid)==0                    /* (1) */
            && (bestJ<0 || (notIndexed&m)!=0                     /* (2) */
                || (bestPlan.plan.wsFlags & WHERE_NOT_FULLSCAN)==0
                || (sWBI.cost.plan.wsFlags & WHERE_NOT_FULLSCAN)!=0)
            && (nUnconstrained==0 || sWBI.pSrc->pIndex==0        /* (3) */
                || NEVER((sWBI.cost.plan.wsFlags & WHERE_NOT_FULLSCAN)!=0))
            && (bestJ<0 || compareCost(&sWBI.cost, &bestPlan))   /* (4) */


        ){
          WHERETRACE(("=== table %d (%s) is best so far\n"
                      "    cost=%.1f, nRow=%.1f, nOBSat=%d, wsFlags=%08x\n",
                      j, sWBI.pSrc->pTab->zName,
                      sWBI.cost.rCost, sWBI.cost.plan.nRow,
                      sWBI.cost.plan.nOBSat, sWBI.cost.plan.wsFlags));
          bestPlan = sWBI.cost;
          bestJ = j;
        }
        if( doNotReorder ) break;
      }
    }
    assert( bestJ>=0 );
    assert( sWBI.notValid & getMask(pMaskSet, pTabList->a[bestJ].iCursor) );
    WHERETRACE(("*** Optimizer selects table %d (%s) for loop %d with:\n"
                "    cost=%.1f, nRow=%.1f, nOBSat=%d, wsFlags=0x%08x\n",
                bestJ, pTabList->a[bestJ].pTab->zName,
                pLevel-pWInfo->a, bestPlan.rCost, bestPlan.plan.nRow,
                bestPlan.plan.nOBSat, bestPlan.plan.wsFlags));



    if( (bestPlan.plan.wsFlags & WHERE_DISTINCT)!=0 ){
      assert( pWInfo->eDistinct==0 );
      pWInfo->eDistinct = WHERE_DISTINCT_ORDERED;
    }
    andFlags &= bestPlan.plan.wsFlags;
    pLevel->plan = bestPlan.plan;
    pLevel->iTabCur = pTabList->a[bestJ].iCursor;

      }
    }
  }
  WHERETRACE(("*** Optimizer Finished ***\n"));
  if( pParse->nErr || db->mallocFailed ){
    goto whereBeginError;
  }
  if( nTabList ){
    pLevel--;
    pWInfo->nOBSat = pLevel->plan.nOBSat;
  }else{
    pWInfo->nOBSat = 0;
  }

  /* If the total query only selects a single row, then the ORDER BY
  ** clause is irrelevant.
  */
  if( (andFlags & WHERE_UNIQUE)!=0 && pOrderBy ){
    assert( nTabList==0 || (pLevel->plan.wsFlags & WHERE_ALL_UNIQUE)!=0 );
    pWInfo->nOBSat = pOrderBy->nExpr;
  }

  /* If the caller is an UPDATE or DELETE statement that is requesting
  ** to use a one-pass algorithm, determine if this is appropriate.
  ** The one-pass algorithm only works if the WHERE clause constraints
  ** the statement to update a single row.

  INSERT INTO x5_rules VALUES(0, 'a', '0.1.2.3.4.5.6.7.8.9.a', 1);
  DROP TABLE x5;
  CREATE VIRTUAL TABLE x5 USING fuzzer(x5_rules);
  SELECT length(word) FROM x5 WHERE word MATCH 'a' LIMIT 50;
} {1 21 41 61 81}

finish_test

do_test 2.1 {
  db eval {
    SELECT a||','||c||','||e||','||g FROM t31, t32, t33, t34
     WHERE c=b AND e=d AND g=f
     ORDER BY +a ASC, +c ASC, +e DESC, +g ASC;
  }
} {1,3,7,10 1,3,7,14 1,3,6,11 1,4,8,12 1,4,8,12 1,4,8,13 1,4,5,9 2,3,7,10 2,3,7,14 2,3,6,11}


finish_test