where.c

最新的sqlite3.6.2源代码

  Index *pProbe;              /* An index we are evaluating */
  int rev;                    /* True to scan in reverse order */
  int flags;                  /* Flags associated with pProbe */
  int nEq;                    /* Number of == or IN constraints */
  int eqTermMask;             /* Mask of valid equality operators */
  double cost;                /* Cost of using pProbe */

  WHERETRACE(("bestIndex: tbl=%s notReady=%llx\n", pSrc->pTab->zName, notReady));
  lowestCost = SQLITE_BIG_DBL;
  pProbe = pSrc->pTab->pIndex;

  /* If the table has no indices and there are no terms in the where
  ** clause that refer to the ROWID, then we will never be able to do
  ** anything other than a full table scan on this table.  We might as
  ** well put it first in the join order.  That way, perhaps it can be
  ** referenced by other tables in the join.
  */
  if( pProbe==0 &&
     findTerm(pWC, iCur, -1, 0, WO_EQ|WO_IN|WO_LT|WO_LE|WO_GT|WO_GE,0)==0 &&
     (pOrderBy==0 || !sortableByRowid(iCur, pOrderBy, pWC->pMaskSet, &rev)) ){
    *pFlags = 0;
    *ppIndex = 0;
    *pnEq = 0;
    return 0.0;
  }

  /* Check for a rowid=EXPR or rowid IN (...) constraints
  */
  pTerm = findTerm(pWC, iCur, -1, notReady, WO_EQ|WO_IN, 0);
  if( pTerm ){
    Expr *pExpr;
    *ppIndex = 0;
    bestFlags = WHERE_ROWID_EQ;
    if( pTerm->eOperator & WO_EQ ){
      /* Rowid== is always the best pick.  Look no further.  Because only
      ** a single row is generated, output is always in sorted order */
      *pFlags = WHERE_ROWID_EQ | WHERE_UNIQUE;
      *pnEq = 1;
      WHERETRACE(("... best is rowid\n"));
      return 0.0;
    }else if( (pExpr = pTerm->pExpr)->pList!=0 ){
      /* Rowid IN (LIST): cost is NlogN where N is the number of list
      ** elements.  */
      lowestCost = pExpr->pList->nExpr;
      lowestCost *= estLog(lowestCost);
    }else{
      /* Rowid IN (SELECT): cost is NlogN where N is the number of rows
      ** in the result of the inner select.  We have no way to estimate
      ** that value so make a wild guess. */
      lowestCost = 200;
    }
    WHERETRACE(("... rowid IN cost: %.9g\n", lowestCost));
  }

  /* Estimate the cost of a table scan.  If we do not know how many
  ** entries are in the table, use 1 million as a guess.
  */
  cost = pProbe ? pProbe->aiRowEst[0] : 1000000;
  WHERETRACE(("... table scan base cost: %.9g\n", cost));
  flags = WHERE_ROWID_RANGE;

  /* Check for constraints on a range of rowids in a table scan.
  */
  pTerm = findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE|WO_GT|WO_GE, 0);
  if( pTerm ){
    if( findTerm(pWC, iCur, -1, notReady, WO_LT|WO_LE, 0) ){
      flags |= WHERE_TOP_LIMIT;
      cost /= 3;  /* Guess that rowid<EXPR eliminates two-thirds or rows */
    }
    if( findTerm(pWC, iCur, -1, notReady, WO_GT|WO_GE, 0) ){
      flags |= WHERE_BTM_LIMIT;
      cost /= 3;  /* Guess that rowid>EXPR eliminates two-thirds of rows */
    }
    WHERETRACE(("... rowid range reduces cost to %.9g\n", cost));
  }else{
    flags = 0;
  }

  /* If the table scan does not satisfy the ORDER BY clause, increase
  ** the cost by NlogN to cover the expense of sorting. */
  if( pOrderBy ){
    if( sortableByRowid(iCur, pOrderBy, pWC->pMaskSet, &rev) ){
      flags |= WHERE_ORDERBY|WHERE_ROWID_RANGE;
      if( rev ){
        flags |= WHERE_REVERSE;
      }
    }else{
      cost += cost*estLog(cost);
      WHERETRACE(("... sorting increases cost to %.9g\n", cost));
    }
  }
  if( cost<lowestCost ){
    lowestCost = cost;
    bestFlags = flags;
  }

  /* If the pSrc table is the right table of a LEFT JOIN then we may not
  ** use an index to satisfy IS NULL constraints on that table.  This is
  ** because columns might end up being NULL if the table does not match -
  ** a circumstance which the index cannot help us discover.  Ticket #2177.
  */
  if( (pSrc->jointype & JT_LEFT)!=0 ){
    eqTermMask = WO_EQ|WO_IN;
  }else{
    eqTermMask = WO_EQ|WO_IN|WO_ISNULL;
  }

  /* Look at each index.
  */
  for(; pProbe; pProbe=pProbe->pNext){
    int i;                       /* Loop counter */
    double inMultiplier = 1;

    WHERETRACE(("... index %s:\n", pProbe->zName));

    /* Count the number of columns in the index that are satisfied
    ** by x=EXPR constraints or x IN (...) constraints.
    */
    flags = 0;
    for(i=0; i<pProbe->nColumn; i++){
      int j = pProbe->aiColumn[i];
      pTerm = findTerm(pWC, iCur, j, notReady, eqTermMask, pProbe);
      if( pTerm==0 ) break;
      flags |= WHERE_COLUMN_EQ;
      if( pTerm->eOperator & WO_IN ){
        Expr *pExpr = pTerm->pExpr;
        flags |= WHERE_COLUMN_IN;
        if( pExpr->pSelect!=0 ){
          inMultiplier *= 25;
        }else if( ALWAYS(pExpr->pList) ){
          inMultiplier *= pExpr->pList->nExpr + 1;
        }
      }
    }
    cost = pProbe->aiRowEst[i] * inMultiplier * estLog(inMultiplier);
    nEq = i;
    if( pProbe->onError!=OE_None && (flags & WHERE_COLUMN_IN)==0
         && nEq==pProbe->nColumn ){
      flags |= WHERE_UNIQUE;
    }
    WHERETRACE(("...... nEq=%d inMult=%.9g cost=%.9g\n",nEq,inMultiplier,cost));

    /* Look for range constraints
    */
    if( nEq<pProbe->nColumn ){
      int j = pProbe->aiColumn[nEq];
      pTerm = findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE|WO_GT|WO_GE, pProbe);
      if( pTerm ){
        flags |= WHERE_COLUMN_RANGE;
        if( findTerm(pWC, iCur, j, notReady, WO_LT|WO_LE, pProbe) ){
          flags |= WHERE_TOP_LIMIT;
          cost /= 3;
        }
        if( findTerm(pWC, iCur, j, notReady, WO_GT|WO_GE, pProbe) ){
          flags |= WHERE_BTM_LIMIT;
          cost /= 3;
        }
        WHERETRACE(("...... range reduces cost to %.9g\n", cost));
      }
    }

    /* Add the additional cost of sorting if that is a factor.
    */
    if( pOrderBy ){
      if( (flags & WHERE_COLUMN_IN)==0 &&
           isSortingIndex(pParse,pWC->pMaskSet,pProbe,iCur,pOrderBy,nEq,&rev) ){
        if( flags==0 ){
          flags = WHERE_COLUMN_RANGE;
        }
        flags |= WHERE_ORDERBY;
        if( rev ){
          flags |= WHERE_REVERSE;
        }
      }else{
        cost += cost*estLog(cost);
        WHERETRACE(("...... orderby increases cost to %.9g\n", cost));
      }
    }

    /* Check to see if we can get away with using just the index without
    ** ever reading the table.  If that is the case, then halve the
    ** cost of this index.
    */
    if( flags && pSrc->colUsed < (((Bitmask)1)<<(BMS-1)) ){
      Bitmask m = pSrc->colUsed;
      int j;
      for(j=0; j<pProbe->nColumn; j++){
        int x = pProbe->aiColumn[j];
        if( x<BMS-1 ){
          m &= ~(((Bitmask)1)<<x);
        }
      }
      if( m==0 ){
        flags |= WHERE_IDX_ONLY;
        cost /= 2;
        WHERETRACE(("...... idx-only reduces cost to %.9g\n", cost));
      }
    }

    /* If this index has achieved the lowest cost so far, then use it.
    */
    if( flags && cost < lowestCost ){
      bestIdx = pProbe;
      lowestCost = cost;
      bestFlags = flags;
      bestNEq = nEq;
    }
  }

  /* Report the best result
  */
  *ppIndex = bestIdx;
  WHERETRACE(("best index is %s, cost=%.9g, flags=%x, nEq=%d\n",
        bestIdx ? bestIdx->zName : "(none)", lowestCost, bestFlags, bestNEq));
  *pFlags = bestFlags | eqTermMask;
  *pnEq = bestNEq;
  return lowestCost;
}


/*
** Disable a term in the WHERE clause.  Except, do not disable the term
** if it controls a LEFT OUTER JOIN and it did not originate in the ON
** or USING clause of that join.
**
** Consider the term t2.z='ok' in the following queries:
**
**   (1)  SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x WHERE t2.z='ok'
**   (2)  SELECT * FROM t1 LEFT JOIN t2 ON t1.a=t2.x AND t2.z='ok'
**   (3)  SELECT * FROM t1, t2 WHERE t1.a=t2.x AND t2.z='ok'
**
** The t2.z='ok' is disabled in the in (2) because it originates
** in the ON clause.  The term is disabled in (3) because it is not part
** of a LEFT OUTER JOIN.  In (1), the term is not disabled.
**
** Disabling a term causes that term to not be tested in the inner loop
** of the join.  Disabling is an optimization.  When terms are satisfied
** by indices, we disable them to prevent redundant tests in the inner
** loop.  We would get the correct results if nothing were ever disabled,
** but joins might run a little slower.  The trick is to disable as much
** as we can without disabling too much.  If we disabled in (1), we'd get
** the wrong answer.  See ticket #813.
*/
static void disableTerm(WhereLevel *pLevel, WhereTerm *pTerm){
  if( pTerm
      && ALWAYS((pTerm->flags & TERM_CODED)==0)
      && (pLevel->iLeftJoin==0 || ExprHasProperty(pTerm->pExpr, EP_FromJoin))
  ){
    pTerm->flags |= TERM_CODED;
    if( pTerm->iParent>=0 ){
      WhereTerm *pOther = &pTerm->pWC->a[pTerm->iParent];
      if( (--pOther->nChild)==0 ){
        disableTerm(pLevel, pOther);
      }
    }
  }
}

/*
** Apply the affinities associated with the first n columns of index
** pIdx to the values in the n registers starting at base.
*/
static void codeApplyAffinity(Parse *pParse, int base, int n, Index *pIdx){
  if( n>0 ){
    Vdbe *v = pParse->pVdbe;
    assert( v!=0 );
    sqlite3VdbeAddOp2(v, OP_Affinity, base, n);
    sqlite3IndexAffinityStr(v, pIdx);
    sqlite3ExprCacheAffinityChange(pParse, base, n);
  }
}


/*
** Generate code for a single equality term of the WHERE clause.  An equality
** term can be either X=expr or X IN (...).   pTerm is the term to be 
** coded.
**
** The current value for the constraint is left in register iReg.
**
** For a constraint of the form X=expr, the expression is evaluated and its
** result is left on the stack.  For constraints of the form X IN (...)
** this routine sets up a loop that will iterate over all values of X.
*/
static int codeEqualityTerm(
  Parse *pParse,      /* The parsing context */
  WhereTerm *pTerm,   /* The term of the WHERE clause to be coded */
  WhereLevel *pLevel, /* When level of the FROM clause we are working on */
  int iTarget         /* Attempt to leave results in this register */
){
  Expr *pX = pTerm->pExpr;
  Vdbe *v = pParse->pVdbe;
  int iReg;                  /* Register holding results */

  if( iTarget<=0 ){
    iReg = iTarget = sqlite3GetTempReg(pParse);
  }
  if( pX->op==TK_EQ ){
    iReg = sqlite3ExprCodeTarget(pParse, pX->pRight, iTarget);
  }else if( pX->op==TK_ISNULL ){
    iReg = iTarget;
    sqlite3VdbeAddOp2(v, OP_Null, 0, iReg);
#ifndef SQLITE_OMIT_SUBQUERY
  }else{
    int eType;
    int iTab;
    struct InLoop *pIn;

    assert( pX->op==TK_IN );
    iReg = iTarget;
    eType = sqlite3FindInIndex(pParse, pX, 0);
    iTab = pX->iTable;
    sqlite3VdbeAddOp2(v, OP_Rewind, iTab, 0);
    VdbeComment((v, "%.*s", pX->span.n, pX->span.z));
    if( pLevel->nIn==0 ){
      pLevel->nxt = sqlite3VdbeMakeLabel(v);
    }
    pLevel->nIn++;
    pLevel->aInLoop = sqlite3DbReallocOrFree(pParse->db, pLevel->aInLoop,
                                    sizeof(pLevel->aInLoop[0])*pLevel->nIn);
    pIn = pLevel->aInLoop;
    if( pIn ){
      pIn += pLevel->nIn - 1;
      pIn->iCur = iTab;
      if( eType==IN_INDEX_ROWID ){
        pIn->topAddr = sqlite3VdbeAddOp2(v, OP_Rowid, iTab, iReg);
      }else{
        pIn->topAddr = sqlite3VdbeAddOp3(v, OP_Column, iTab, 0, iReg);
      }
      sqlite3VdbeAddOp1(v, OP_IsNull, iReg);
    }else{
      pLevel->nIn = 0;
    }
#endif
  }
  disableTerm(pLevel, pTerm);
  return iReg;
}

/*
** Generate code that will evaluate all == and IN constraints for an
** index.  The values for all constraints are left on the stack.
**
** For example, consider table t1(a,b,c,d,e,f) with index i1(a,b,c).
** Suppose the WHERE clause is this:  a==5 AND b IN (1,2,3) AND c>5 AND c<10
** The index has as many as three equality constraints, but in this
** example, the third "c" value is an inequality.  So only two 
** constraints are coded.  This routine will generate code to evaluate
** a==5 and b IN (1,2,3).  The current values for a and b will be left
** on the stack - a is the deepest and b the shallowest.
**
** In the example above nEq==2.  But this subroutine works for any value
** of nEq including 0.  If nEq==0, this routine is nearly a no-op.
** The only thing it does is allocate the pLevel->iMem memory cell.
**
** This routine always allocates at least one memory cell and puts
** the address of that memory cell in pLevel->iMem.  The code that
** calls this routine will use pLevel->iMem to store the termination
** key value of the loop.  If one or more IN operators appear, then
** this routine allocates an additional nEq memory cells for internal
** use.
*/
static int codeAllEqualityTerms(
  Parse *pParse,        /* Parsing context */
  WhereLevel *pLevel,   /* Which nested loop of the FROM we are coding */
  WhereClause *pWC,     /* The WHERE clause */
  Bitmask notReady,     /* Which parts of FROM have not yet been coded */
  int nExtraReg         /* Number of extra registers to allocate */
){
  int nEq = pLevel->nEq;        /* The number of == or IN constraints to code */
  Vdbe *v = pParse->pVdbe;      /* The virtual machine under construction */
  Index *pIdx =