SQLite

** the page, 1 means the second cell, and so forth) return a pointer
** to the cell content.
**
** This routine works only for pages that do not contain overflow cells.
*/
#define findCell(P,I) \
  ((P)->aData + ((P)->maskPage & get2byte(&(P)->aCellIdx[2*(I)])))
#define findCellv2(D,M,O,I) (D+(M&get2byte(D+(O+2*(I)))))


/*
** This a more complex version of findCell() that works for
** pages that do contain overflow cells.
*/
static u8 *findOverflowCell(MemPage *pPage, int iCell){
  int i;
  assert( sqlite3_mutex_held(pPage->pBt->mutex) );
  for(i=pPage->nOverflow-1; i>=0; i--){
    int k;
    k = pPage->aiOvfl[i];
    if( k<=iCell ){
      if( k==iCell ){
        return pPage->apOvfl[i];
      }
      iCell--;
    }
  }
  return findCell(pPage, iCell);
}

/*
** This is common tail processing for btreeParseCellPtr() and
** btreeParseCellPtrIndex() for the case when the cell does not fit entirely
** on a single B-tree page.  Make necessary adjustments to the CellInfo
** structure.
*/

    **
    ** Check to see if iFreeBlk should be coalesced onto the end of iStart.
    */
    if( iFreeBlk && iEnd+3>=iFreeBlk ){
      nFrag = iFreeBlk - iEnd;
      if( iEnd>iFreeBlk ) return SQLITE_CORRUPT_BKPT;
      iEnd = iFreeBlk + get2byte(&data[iFreeBlk+2]);
      if( iEnd > pPage->pBt->usableSize ) return SQLITE_CORRUPT_BKPT;
      if( NEVER(iEnd > pPage->pBt->usableSize) ) return SQLITE_CORRUPT_BKPT;
      iSize = iEnd - iStart;
      iFreeBlk = get2byte(&data[iFreeBlk]);
    }
  
    /* If iPtr is another freeblock (that is, if iPtr is not the freelist
    ** pointer in the page header) then check to see if iStart should be
    ** coalesced onto the end of iPtr.

    if( iChild ){
      put4byte(pCell, iChild);
    }
    j = pPage->nOverflow++;
    assert( j<(int)(sizeof(pPage->apOvfl)/sizeof(pPage->apOvfl[0])) );
    pPage->apOvfl[j] = pCell;
    pPage->aiOvfl[j] = (u16)i;

    /* When multiple overflows occur, they are always sequential and in
    ** sorted order.  This invariants arise because multiple overflows can
    ** only occur when inserting divider cells into the parent page during
    ** balancing, and the dividers are adjacent and sorted.
    */
    assert( j==0 || pPage->aiOvfl[j-1]<(u16)i ); /* Overflows in sorted order */
    assert( j==0 || i==pPage->aiOvfl[j-1]+1 );   /* Overflows are sequential */
  }else{
    int rc = sqlite3PagerWrite(pPage->pDbPage);
    if( rc!=SQLITE_OK ){
      *pRC = rc;
      return;
    }
    assert( sqlite3PagerIswriteable(pPage->pDbPage) );

      ** the entry for the overflow page into the pointer map.
      */
      ptrmapPutOvflPtr(pPage, pCell, pRC);
    }
#endif
  }
}

/*
** A CellArray object contains a cache of pointers and sizes for a
** consecutive sequence of cells that might be held multiple pages.
*/
typedef struct CellArray CellArray;
struct CellArray {
  int nCell;              /* Number of cells in apCell[] */
  MemPage *pRef;          /* Reference page */
  u8 **apCell;            /* All cells begin balanced */
  u16 *szCell;            /* Local size of all cells in apCell[] */
};

/*
** Make sure the cell sizes at idx, idx+1, ..., idx+N-1 have been
** computed.
*/
static void populateCellCache(CellArray *p, int idx, int N){
  assert( idx>=0 && idx+N<=p->nCell );
  while( N>0 ){
    assert( p->apCell[idx]!=0 );
    if( p->szCell[idx]==0 ){
      p->szCell[idx] = p->pRef->xCellSize(p->pRef, p->apCell[idx]);
    }else{
      assert( CORRUPT_DB ||
              p->szCell[idx]==p->pRef->xCellSize(p->pRef, p->apCell[idx]) );
    }
    idx++;
    N--;
  }
}

/*
** Return the size of the Nth element of the cell array
*/
static SQLITE_NOINLINE u16 computeCellSize(CellArray *p, int N){
  assert( N>=0 && N<p->nCell );
  assert( p->szCell[N]==0 );
  p->szCell[N] = p->pRef->xCellSize(p->pRef, p->apCell[N]);
  return p->szCell[N];
}
static u16 cachedCellSize(CellArray *p, int N){
  assert( N>=0 && N<p->nCell );
  if( p->szCell[N] ) return p->szCell[N];
  return computeCellSize(p, N);
}

/*
** Array apCell[] contains pointers to nCell b-tree page cells. The 
** szCell[] array contains the size in bytes of each cell. This function
** replaces the current contents of page pPg with the contents of the cell
** array.
**
** Some of the cells in apCell[] may currently be stored in pPg. This
** function works around problems caused by this by making a copy of any 
** such cells before overwriting the page data.
**
** The MemPage.nFree field is invalidated by this function. It is the 
** responsibility of the caller to set it correctly.
*/
static void rebuildPage(
static int rebuildPage(
  MemPage *pPg,                   /* Edit this page */
  int nCell,                      /* Final number of cells on page */
  u8 **apCell,                    /* Array of cells */
  u16 *szCell                     /* Array of cell sizes */
){
  const int hdr = pPg->hdrOffset;          /* Offset of header on pPg */
  u8 * const aData = pPg->aData;           /* Pointer to data for pPg */

  pData = pEnd;
  for(i=0; i<nCell; i++){
    u8 *pCell = apCell[i];
    if( pCell>aData && pCell<pEnd ){
      pCell = &pTmp[pCell - aData];
    }
    pData -= szCell[i];
    memcpy(pData, pCell, szCell[i]);
    put2byte(pCellptr, (pData - aData));
    pCellptr += 2;
    if( pData < pCellptr ) return SQLITE_CORRUPT_BKPT;
    memcpy(pData, pCell, szCell[i]);
    assert( szCell[i]==pPg->xCellSize(pPg, pCell) || CORRUPT_DB );
    testcase( szCell[i]==pPg->xCellSize(pPg,pCell) );
    testcase( szCell[i]!=pPg->xCellSize(pPg,pCell) );
  }

  /* The pPg->nFree field is now set incorrectly. The caller will fix it. */
  pPg->nCell = nCell;
  pPg->nOverflow = 0;

  put2byte(&aData[hdr+1], 0);
  put2byte(&aData[hdr+3], pPg->nCell);
  put2byte(&aData[hdr+5], pData - aData);
  aData[hdr+7] = 0x00;
  return SQLITE_OK;
}

/*
** Array apCell[] contains nCell pointers to b-tree cells. Array szCell
** contains the size in bytes of each such cell. This function attempts to 
** add the cells stored in the array to page pPg. If it cannot (because 
** the page needs to be defragmented before the cells will fit), non-zero

** cells in apCell[], then the cells do not fit and non-zero is returned.
*/
static int pageInsertArray(
  MemPage *pPg,                   /* Page to add cells to */
  u8 *pBegin,                     /* End of cell-pointer array */
  u8 **ppData,                    /* IN/OUT: Page content -area pointer */
  u8 *pCellptr,                   /* Pointer to cell-pointer area */
  int iFirst,                     /* Index of first cell to add */
  int nCell,                      /* Number of cells to add to pPg */
  u8 **apCell,                    /* Array of cells */
  CellArray *pCArray              /* Array of cells */
  u16 *szCell                     /* Array of cell sizes */
){
  int i;
  u8 *aData = pPg->aData;
  u8 *pData = *ppData;
  const int bFreelist = aData[1] || aData[2];
  int iEnd = iFirst + nCell;
  assert( CORRUPT_DB || pPg->hdrOffset==0 );    /* Never called on page 1 */
  for(i=0; i<nCell; i++){
    int sz = szCell[i];
  for(i=iFirst; i<iEnd; i++){
    int sz, rc;
    int rc;
    u8 *pSlot;
    sz = cachedCellSize(pCArray, i);
    if( bFreelist==0 || (pSlot = pageFindSlot(pPg, sz, &rc, 0))==0 ){
      pData -= sz;
      if( pData<pBegin ) return 1;
      pSlot = pData;
    }
    memcpy(pSlot, apCell[i], sz);
    memcpy(pSlot, pCArray->apCell[i], sz);
    put2byte(pCellptr, (pSlot - aData));
    pCellptr += 2;
  }
  *ppData = pData;
  return 0;
}

/*
** Array apCell[] contains nCell pointers to b-tree cells. Array szCell 
** contains the size in bytes of each such cell. This function adds the
** space associated with each cell in the array that is currently stored 
** within the body of pPg to the pPg free-list. The cell-pointers and other
** fields of the page are not updated.
**
** This function returns the total number of cells added to the free-list.
*/
static int pageFreeArray(
  MemPage *pPg,                   /* Page to edit */
  int iFirst,                     /* First cell to delete */
  int nCell,                      /* Cells to delete */
  u8 **apCell,                    /* Array of cells */
  CellArray *pCArray              /* Array of cells */
  u16 *szCell                     /* Array of cell sizes */
){
  u8 * const aData = pPg->aData;
  u8 * const pEnd = &aData[pPg->pBt->usableSize];
  u8 * const pStart = &aData[pPg->hdrOffset + 8 + pPg->childPtrSize];
  int nRet = 0;
  int i;
  int iEnd = iFirst + nCell;
  u8 *pFree = 0;
  int szFree = 0;

  for(i=0; i<nCell; i++){
    u8 *pCell = apCell[i];
  for(i=iFirst; i<iEnd; i++){
    u8 *pCell = pCArray->apCell[i];
    if( pCell>=pStart && pCell<pEnd ){
      int sz = szCell[i];
      int sz;
      /* No need to use cachedCellSize() here.  The sizes of all cells that
      ** are to be freed have already been computing while deciding which
      ** cells need freeing */
      sz = pCArray->szCell[i];  assert( sz>0 );
      if( pFree!=(pCell + sz) ){
        if( pFree ){
          assert( pFree>aData && (pFree - aData)<65536 );
          freeSpace(pPg, (u16)(pFree - aData), szFree);
        }
        pFree = pCell;
        szFree = sz;

**
** This routine makes the necessary adjustments to pPg so that it contains
** the correct cells after being balanced.
**
** The pPg->nFree field is invalid when this function returns. It is the
** responsibility of the caller to set it correctly.
*/
static void editPage(
static int editPage(
  MemPage *pPg,                   /* Edit this page */
  int iOld,                       /* Index of first cell currently on page */
  int iNew,                       /* Index of new first cell on page */
  int nNew,                       /* Final number of cells on page */
  u8 **apCell,                    /* Array of cells */
  CellArray *pCArray              /* Array of cells and sizes */
  u16 *szCell                     /* Array of cell sizes */
){
  u8 * const aData = pPg->aData;
  const int hdr = pPg->hdrOffset;
  u8 *pBegin = &pPg->aCellIdx[nNew * 2];
  int nCell = pPg->nCell;       /* Cells stored on pPg */
  u8 *pData;
  u8 *pCellptr;
  int i;
  int iOldEnd = iOld + pPg->nCell + pPg->nOverflow;
  int iNewEnd = iNew + nNew;

#ifdef SQLITE_DEBUG
  u8 *pTmp = sqlite3PagerTempSpace(pPg->pBt->pPager);
  memcpy(pTmp, aData, pPg->pBt->usableSize);
#endif

  /* Remove cells from the start and end of the page */
  if( iOld<iNew ){
    int nShift = pageFreeArray(
    int nShift = pageFreeArray(pPg, iOld, iNew-iOld, pCArray);
        pPg, iNew-iOld, &apCell[iOld], &szCell[iOld]
    );
    memmove(pPg->aCellIdx, &pPg->aCellIdx[nShift*2], nCell*2);
    nCell -= nShift;
  }
  if( iNewEnd < iOldEnd ){
    nCell -= pageFreeArray(
    nCell -= pageFreeArray(pPg, iNewEnd, iOldEnd - iNewEnd, pCArray);
        pPg, iOldEnd-iNewEnd, &apCell[iNewEnd], &szCell[iNewEnd]
    );
  }

  pData = &aData[get2byteNotZero(&aData[hdr+5])];
  if( pData<pBegin ) goto editpage_fail;

  /* Add cells to the start of the page */
  if( iNew<iOld ){
    int nAdd = MIN(nNew,iOld-iNew);
    assert( (iOld-iNew)<nNew || nCell==0 || CORRUPT_DB );
    pCellptr = pPg->aCellIdx;
    memmove(&pCellptr[nAdd*2], pCellptr, nCell*2);
    if( pageInsertArray(
          pPg, pBegin, &pData, pCellptr,
          nAdd, &apCell[iNew], &szCell[iNew]
          iNew, nAdd, pCArray
    ) ) goto editpage_fail;
    nCell += nAdd;
  }

  /* Add any overflow cells */
  for(i=0; i<pPg->nOverflow; i++){
    int iCell = (iOld + pPg->aiOvfl[i]) - iNew;
    if( iCell>=0 && iCell<nNew ){
      pCellptr = &pPg->aCellIdx[iCell * 2];
      memmove(&pCellptr[2], pCellptr, (nCell - iCell) * 2);
      nCell++;
      if( pageInsertArray(
            pPg, pBegin, &pData, pCellptr,
            1, &apCell[iCell + iNew], &szCell[iCell + iNew]
            iCell+iNew, 1, pCArray
      ) ) goto editpage_fail;
    }
  }

  /* Append cells to the end of the page */
  pCellptr = &pPg->aCellIdx[nCell*2];
  if( pageInsertArray(
        pPg, pBegin, &pData, pCellptr,
        nNew-nCell, &apCell[iNew+nCell], &szCell[iNew+nCell]
        iNew+nCell, nNew-nCell, pCArray
  ) ) goto editpage_fail;

  pPg->nCell = nNew;
  pPg->nOverflow = 0;

  put2byte(&aData[hdr+3], pPg->nCell);
  put2byte(&aData[hdr+5], pData - aData);

#ifdef SQLITE_DEBUG
  for(i=0; i<nNew && !CORRUPT_DB; i++){
    u8 *pCell = apCell[i+iNew];
    u8 *pCell = pCArray->apCell[i+iNew];
    int iOff = get2byte(&pPg->aCellIdx[i*2]);
    if( pCell>=aData && pCell<&aData[pPg->pBt->usableSize] ){
      pCell = &pTmp[pCell - aData];
    }
    assert( 0==memcmp(pCell, &aData[iOff], szCell[i+iNew]) );
    assert( 0==memcmp(pCell, &aData[iOff],
            pCArray->pRef->xCellSize(pCArray->pRef, pCArray->apCell[i+iNew])) );
  }
#endif

  return;
  return SQLITE_OK;
 editpage_fail:
  /* Unable to edit this page. Rebuild it from scratch instead. */
  populateCellCache(pCArray, iNew, nNew);
  rebuildPage(pPg, nNew, &apCell[iNew], &szCell[iNew]);
  return rebuildPage(pPg, nNew, &pCArray->apCell[iNew], &pCArray->szCell[iNew]);
}

/*
** The following parameters determine how many adjacent pages get involved
** in a balancing operation.  NN is the number of neighbors on either side
** of the page that participate in the balancing operation.  NB is the
** total number of pages that participate, including the target page and

    u8 *pCell = pPage->apOvfl[0];
    u16 szCell = pPage->xCellSize(pPage, pCell);
    u8 *pStop;

    assert( sqlite3PagerIswriteable(pNew->pDbPage) );
    assert( pPage->aData[0]==(PTF_INTKEY|PTF_LEAFDATA|PTF_LEAF) );
    zeroPage(pNew, PTF_INTKEY|PTF_LEAFDATA|PTF_LEAF);
    rebuildPage(pNew, 1, &pCell, &szCell);
    rc = rebuildPage(pNew, 1, &pCell, &szCell);
    if( NEVER(rc) ) return rc;
    pNew->nFree = pBt->usableSize - pNew->cellOffset - 2 - szCell;

    /* If this is an auto-vacuum database, update the pointer map
    ** with entries for the new page, and any pointer from the 
    ** cell on the page to an overflow page. If either of these
    ** operations fails, the return code is set, but the contents
    ** of the parent page are still manipulated by thh code below.

  MemPage *pParent,               /* Parent page of siblings being balanced */
  int iParentIdx,                 /* Index of "the page" in pParent */
  u8 *aOvflSpace,                 /* page-size bytes of space for parent ovfl */
  int isRoot,                     /* True if pParent is a root-page */
  int bBulk                       /* True if this call is part of a bulk load */
){
  BtShared *pBt;               /* The whole database */
  int nCell = 0;               /* Number of cells in apCell[] */
  int nMaxCells = 0;           /* Allocated size of apCell, szCell, aFrom. */
  int nNew = 0;                /* Number of pages in apNew[] */
  int nOld;                    /* Number of pages in apOld[] */
  int i, j, k;                 /* Loop counters */
  int nxDiv;                   /* Next divider slot in pParent->aCell[] */
  int rc = SQLITE_OK;          /* The return code */
  u16 leafCorrection;          /* 4 if pPage is a leaf.  0 if not */
  int leafData;                /* True if pPage is a leaf of a LEAFDATA tree */
  int usableSpace;             /* Bytes in pPage beyond the header */
  int pageFlags;               /* Value of pPage->aData[0] */
  int subtotal;                /* Subtotal of bytes in cells on one page */
  int iSpace1 = 0;             /* First unused byte of aSpace1[] */
  int iOvflSpace = 0;          /* First unused byte of aOvflSpace[] */
  int szScratch;               /* Size of scratch memory requested */
  MemPage *apOld[NB];          /* pPage and up to two siblings */
  MemPage *apNew[NB+2];        /* pPage and up to NB siblings after balancing */
  u8 *pRight;                  /* Location in parent of right-sibling pointer */
  u8 *apDiv[NB-1];             /* Divider cells in pParent */
  int cntNew[NB+2];            /* Index in aCell[] of cell after i-th page */
  int cntOld[NB+2];            /* Old index in aCell[] after i-th page */
  int cntNew[NB+2];            /* Index in b.paCell[] of cell after i-th page */
  int cntOld[NB+2];            /* Old index in b.apCell[] */
  int szNew[NB+2];             /* Combined size of cells placed on i-th page */
  u8 **apCell = 0;             /* All cells begin balanced */
  u16 *szCell;                 /* Local size of all cells in apCell[] */
  u8 *aSpace1;                 /* Space for copies of dividers cells */
  Pgno pgno;                   /* Temp var to store a page number in */
  u8 abDone[NB+2];             /* True after i'th new page is populated */
  Pgno aPgno[NB+2];            /* Page numbers of new pages before shuffling */
  Pgno aPgOrder[NB+2];         /* Copy of aPgno[] used for sorting pages */
  u16 aPgFlags[NB+2];          /* flags field of new pages before shuffling */
  CellArray b;                  /* Parsed information on cells being balanced */

  memset(abDone, 0, sizeof(abDone));
  b.nCell = 0;
  b.apCell = 0;
  pBt = pParent->pBt;
  assert( sqlite3_mutex_held(pBt->mutex) );
  assert( sqlite3PagerIswriteable(pParent->pDbPage) );

#if 0
  TRACE(("BALANCE: begin page %d child of %d\n", pPage->pgno, pParent->pgno));
#endif

  ** alignment */
  nMaxCells = (nMaxCells + 3)&~3;

  /*
  ** Allocate space for memory structures
  */
  szScratch =
       nMaxCells*sizeof(u8*)                       /* apCell */
     + nMaxCells*sizeof(u16)                       /* szCell */
       nMaxCells*sizeof(u8*)                       /* b.apCell */
     + nMaxCells*sizeof(u16)                       /* b.szCell */
     + pBt->pageSize;                              /* aSpace1 */

  /* EVIDENCE-OF: R-28375-38319 SQLite will never request a scratch buffer
  ** that is more than 6 times the database page size. */
  assert( szScratch<=6*(int)pBt->pageSize );
  apCell = sqlite3ScratchMalloc( szScratch ); 
  if( apCell==0 ){
  b.apCell = sqlite3ScratchMalloc( szScratch ); 
  if( b.apCell==0 ){
    rc = SQLITE_NOMEM;
    goto balance_cleanup;
  }
  szCell = (u16*)&apCell[nMaxCells];
  aSpace1 = (u8*)&szCell[nMaxCells];
  b.szCell = (u16*)&b.apCell[nMaxCells];
  aSpace1 = (u8*)&b.szCell[nMaxCells];
  assert( EIGHT_BYTE_ALIGNMENT(aSpace1) );

  /*
  ** Load pointers to all cells on sibling pages and the divider cells
  ** into the local apCell[] array.  Make copies of the divider cells
  ** into the local b.apCell[] array.  Make copies of the divider cells
  ** into space obtained from aSpace1[]. The divider cells have already
  ** been removed from pParent.
  **
  ** If the siblings are on leaf pages, then the child pointers of the
  ** divider cells are stripped from the cells before they are copied
  ** into aSpace1[].  In this way, all cells in apCell[] are without
  ** into aSpace1[].  In this way, all cells in b.apCell[] are without
  ** child pointers.  If siblings are not leaves, then all cell in
  ** apCell[] include child pointers.  Either way, all cells in apCell[]
  ** b.apCell[] include child pointers.  Either way, all cells in b.apCell[]
  ** are alike.
  **
  ** leafCorrection:  4 if pPage is a leaf.  0 if pPage is not a leaf.
  **       leafData:  1 if pPage holds key+data and pParent holds only keys.
  */
  b.pRef = apOld[0];
  leafCorrection = apOld[0]->leaf*4;
  leafData = apOld[0]->intKeyLeaf;
  leafCorrection = b.pRef->leaf*4;
  leafData = b.pRef->intKeyLeaf;
  for(i=0; i<nOld; i++){
    int limit;
    MemPage *pOld = apOld[i];
    int limit = pOld->nCell;
    u8 *aData = pOld->aData;
    u16 maskPage = pOld->maskPage;
    u8 *piCell = aData + pOld->cellOffset;
    u8 *piEnd;

    /* Verify that all sibling pages are of the same "type" (table-leaf,
    ** table-interior, index-leaf, or index-interior).
    */
    if( pOld->aData[0]!=apOld[0]->aData[0] ){
      rc = SQLITE_CORRUPT_BKPT;
      goto balance_cleanup;
    }

    /* Load b.apCell[] with pointers to all cells in pOld.  If pOld
    limit = pOld->nCell+pOld->nOverflow;
    ** constains overflow cells, include them in the b.apCell[] array
    ** in the correct spot.
    **
    ** Note that when there are multiple overflow cells, it is always the
    ** case that they are sequential and adjacent.  This invariant arises
    ** because multiple overflows can only occurs when inserting divider
    ** cells into a parent on a prior balance, and divider cells are always
    ** adjacent and are inserted in order.  There is an assert() tagged
    ** with "NOTE 1" in the overflow cell insertion loop to prove this
    ** invariant.
    **
    ** This must be done in advance.  Once the balance starts, the cell
    ** offset section of the btree page will be overwritten and we will no
    ** long be able to find the cells if a pointer to each cell is not saved
    ** first.
    */
    memset(&b.szCell[b.nCell], 0, sizeof(b.szCell[0])*limit);
    if( pOld->nOverflow>0 ){
      memset(&b.szCell[b.nCell+limit], 0, sizeof(b.szCell[0])*pOld->nOverflow);
      limit = pOld->aiOvfl[0];
      for(j=0; j<limit; j++){
        b.apCell[b.nCell] = aData + (maskPage & get2byte(piCell));
        piCell += 2;
        assert( nCell<nMaxCells );
        apCell[nCell] = findOverflowCell(pOld, j);
        szCell[nCell] = pOld->xCellSize(pOld, apCell[nCell]);
        nCell++;
        b.nCell++;
      }
      for(k=0; k<pOld->nOverflow; k++){
        assert( k==0 || pOld->aiOvfl[k-1]+1==pOld->aiOvfl[k] );/* NOTE 1 */
        b.apCell[b.nCell] = pOld->apOvfl[k];
        b.nCell++;
      }
    }else{
      u8 *aData = pOld->aData;
      limit = pOld->nCell - pOld->aiOvfl[0];
      u16 maskPage = pOld->maskPage;
      u16 cellOffset = pOld->cellOffset;
    }
    piEnd = aData + pOld->cellOffset + 2*pOld->nCell;
      for(j=0; j<limit; j++){
        assert( nCell<nMaxCells );
        apCell[nCell] = findCellv2(aData, maskPage, cellOffset, j);
        szCell[nCell] = pOld->xCellSize(pOld, apCell[nCell]);
        nCell++;
      }
    }       
    cntOld[i] = nCell;
    while( piCell<piEnd ){
      assert( b.nCell<nMaxCells );
      b.apCell[b.nCell] = aData + (maskPage & get2byte(piCell));
      piCell += 2;
      b.nCell++;
    }

    cntOld[i] = b.nCell;
    if( i<nOld-1 && !leafData){
      u16 sz = (u16)szNew[i];
      u8 *pTemp;
      assert( nCell<nMaxCells );
      szCell[nCell] = sz;
      assert( b.nCell<nMaxCells );
      b.szCell[b.nCell] = sz;
      pTemp = &aSpace1[iSpace1];
      iSpace1 += sz;
      assert( sz<=pBt->maxLocal+23 );
      assert( iSpace1 <= (int)pBt->pageSize );
      memcpy(pTemp, apDiv[i], sz);
      apCell[nCell] = pTemp+leafCorrection;
      b.apCell[b.nCell] = pTemp+leafCorrection;
      assert( leafCorrection==0 || leafCorrection==4 );
      szCell[nCell] = szCell[nCell] - leafCorrection;
      b.szCell[b.nCell] = b.szCell[b.nCell] - leafCorrection;
      if( !pOld->leaf ){
        assert( leafCorrection==0 );
        assert( pOld->hdrOffset==0 );
        /* The right pointer of the child page pOld becomes the left
        ** pointer of the divider cell */
        memcpy(apCell[nCell], &pOld->aData[8], 4);
        memcpy(b.apCell[b.nCell], &pOld->aData[8], 4);
      }else{
        assert( leafCorrection==4 );
        while( szCell[nCell]<4 ){
        while( b.szCell[b.nCell]<4 ){
          /* Do not allow any cells smaller than 4 bytes. If a smaller cell
          ** does exist, pad it with 0x00 bytes. */
          assert( szCell[nCell]==3 || CORRUPT_DB );
          assert( apCell[nCell]==&aSpace1[iSpace1-3] || CORRUPT_DB );
          assert( b.szCell[b.nCell]==3 || CORRUPT_DB );
          assert( b.apCell[b.nCell]==&aSpace1[iSpace1-3] || CORRUPT_DB );
          aSpace1[iSpace1++] = 0x00;
          szCell[nCell]++;
          b.szCell[b.nCell]++;
        }
      }
      nCell++;
      b.nCell++;
    }
  }

  /*
  ** Figure out the number of pages needed to hold all nCell cells.
  ** Figure out the number of pages needed to hold all b.nCell cells.
  ** Store this number in "k".  Also compute szNew[] which is the total
  ** size of all cells on the i-th page and cntNew[] which is the index
  ** in apCell[] of the cell that divides page i from page i+1.  
  ** cntNew[k] should equal nCell.
  ** in b.apCell[] of the cell that divides page i from page i+1.  
  ** cntNew[k] should equal b.nCell.
  **
  ** Values computed by this block:
  **
  **           k: The total number of sibling pages
  **    szNew[i]: Spaced used on the i-th sibling page.
  **   cntNew[i]: Index in apCell[] and szCell[] for the first cell to
  **   cntNew[i]: Index in b.apCell[] and b.szCell[] for the first cell to
  **              the right of the i-th sibling page.
  ** usableSpace: Number of bytes of space available on each sibling.
  ** 
  */
  usableSpace = pBt->usableSize - 12 + leafCorrection;
  for(subtotal=k=i=0; i<nCell; i++){
    assert( i<nMaxCells );
    subtotal += szCell[i] + 2;
    if( subtotal > usableSpace ){
      szNew[k] = subtotal - szCell[i] - 2;
      cntNew[k] = i;
      if( leafData ){ i--; }
      subtotal = 0;
  for(i=0; i<nOld; i++){
    MemPage *p = apOld[i];
    szNew[i] = usableSpace - p->nFree;
    if( szNew[i]<0 ){ rc = SQLITE_CORRUPT_BKPT; goto balance_cleanup; }
    for(j=0; j<p->nOverflow; j++){
      szNew[i] += 2 + p->xCellSize(p, p->apOvfl[j]);
    }
    cntNew[i] = cntOld[i];
  }
  k = nOld;
  for(i=0; i<k; i++){
    int sz;
    while( szNew[i]>usableSpace ){
      if( i+1>=k ){
        k = i+2;
        if( k>NB+2 ){ rc = SQLITE_CORRUPT_BKPT; goto balance_cleanup; }
        szNew[k-1] = 0;
        cntNew[k-1] = b.nCell;
      }
      sz = 2 + cachedCellSize(&b, cntNew[i]-1);
      szNew[i] -= sz;
      if( !leafData ){
        if( cntNew[i]<b.nCell ){
          sz = 2 + cachedCellSize(&b, cntNew[i]);
        }else{
          sz = 0;
      k++;
      if( k>NB+1 ){ rc = SQLITE_CORRUPT_BKPT; goto balance_cleanup; }
    }
  }
  szNew[k] = subtotal;
  cntNew[k] = nCell;
  k++;
        }
      }
      szNew[i+1] += sz;
      cntNew[i]--;
    }
    while( cntNew[i]<b.nCell ){
      sz = 2 + cachedCellSize(&b, cntNew[i]);
      if( szNew[i]+sz>usableSpace ) break;
      szNew[i] += sz;
      cntNew[i]++;
      if( !leafData ){
        if( cntNew[i]<b.nCell ){
          sz = 2 + cachedCellSize(&b, cntNew[i]);
        }else{
          sz = 0;
        }
      }
      szNew[i+1] -= sz;
    }
    if( cntNew[i]>=b.nCell ){
      k = i+1;
    }else if( cntNew[i] - (i>0 ? cntNew[i-1] : 0) <= 0 ){
      rc = SQLITE_CORRUPT_BKPT;
      goto balance_cleanup;
    }
  }

  /*
  ** The packing computed by the previous block is biased toward the siblings
  ** on the left side (siblings with smaller keys). The left siblings are
  ** always nearly full, while the right-most sibling might be nearly empty.
  ** The next block of code attempts to adjust the packing of siblings to
  ** get a better balance.
  **
  ** This adjustment is more than an optimization.  The packing above might
  ** be so out of balance as to be illegal.  For example, the right-most
  ** sibling might be completely empty.  This adjustment is not optional.
  */
  for(i=k-1; i>0; i--){
    int szRight = szNew[i];  /* Size of sibling on the right */
    int szLeft = szNew[i-1]; /* Size of sibling on the left */
    int r;              /* Index of right-most cell in left sibling */
    int d;              /* Index of first cell to the left of right sibling */

    r = cntNew[i-1] - 1;
    d = r + 1 - leafData;
    (void)cachedCellSize(&b, d);
    while(1){
    assert( d<nMaxCells );
    assert( r<nMaxCells );
    while( szRight==0 
       || (!bBulk && szRight+szCell[d]+2<=szLeft-(szCell[r]+2)) 
      assert( d<nMaxCells );
      assert( r<nMaxCells );
      (void)cachedCellSize(&b, r);
      if( szRight!=0
       && (bBulk || szRight+b.szCell[d]+2 > szLeft-(b.szCell[r]+2)) ){
    ){
      szRight += szCell[d] + 2;
      szLeft -= szCell[r] + 2;
      cntNew[i-1]--;
      r = cntNew[i-1] - 1;
      d = r + 1 - leafData;
        break;
      }
      szRight += b.szCell[d] + 2;
      szLeft -= b.szCell[r] + 2;
      cntNew[i-1] = r;
      if( cntNew[i-1] <= 0 ){
        rc = SQLITE_CORRUPT_BKPT;
        goto balance_cleanup;
      }
      r--;
      d--;
    }
    szNew[i] = szRight;
    szNew[i-1] = szLeft;
  }

  /* Sanity check:  For a non-corrupt database file one of the follwing
  ** must be true:

    }else{
      assert( i>0 );
      rc = allocateBtreePage(pBt, &pNew, &pgno, (bBulk ? 1 : pgno), 0);
      if( rc ) goto balance_cleanup;
      zeroPage(pNew, pageFlags);
      apNew[i] = pNew;
      nNew++;
      cntOld[i] = nCell;
      cntOld[i] = b.nCell;

      /* Set the pointer-map entry for the new sibling page. */
      if( ISAUTOVACUUM ){
        ptrmapPut(pBt, pNew->pgno, PTRMAP_BTREE, pParent->pgno, &rc);
        if( rc!=SQLITE_OK ){
          goto balance_cleanup;
        }

    MemPage *pNew = apNew[0];
    u8 *aOld = pNew->aData;
    int cntOldNext = pNew->nCell + pNew->nOverflow;
    int usableSize = pBt->usableSize;
    int iNew = 0;
    int iOld = 0;

    for(i=0; i<nCell; i++){
      u8 *pCell = apCell[i];
    for(i=0; i<b.nCell; i++){
      u8 *pCell = b.apCell[i];
      if( i==cntOldNext ){
        MemPage *pOld = (++iOld)<nNew ? apNew[iOld] : apOld[iOld];
        cntOldNext += pOld->nCell + pOld->nOverflow + !leafData;
        aOld = pOld->aData;
      }
      if( i==cntNew[iNew] ){
        pNew = apNew[++iNew];

       || pNew->pgno!=aPgno[iOld]
       || pCell<aOld
       || pCell>=&aOld[usableSize]
      ){
        if( !leafCorrection ){
          ptrmapPut(pBt, get4byte(pCell), PTRMAP_BTREE, pNew->pgno, &rc);
        }
        if( szCell[i]>pNew->minLocal ){
        if( cachedCellSize(&b,i)>pNew->minLocal ){
          ptrmapPutOvflPtr(pNew, pCell, &rc);
        }
        if( rc ) goto balance_cleanup;
      }
    }
  }

  /* Insert new divider cells into pParent. */
  for(i=0; i<nNew-1; i++){
    u8 *pCell;
    u8 *pTemp;
    int sz;
    MemPage *pNew = apNew[i];
    j = cntNew[i];

    assert( j<nMaxCells );
    assert( b.apCell[j]!=0 );
    pCell = apCell[j];
    sz = szCell[j] + leafCorrection;
    pCell = b.apCell[j];
    sz = b.szCell[j] + leafCorrection;
    pTemp = &aOvflSpace[iOvflSpace];
    if( !pNew->leaf ){
      memcpy(&pNew->aData[8], pCell, 4);
    }else if( leafData ){
      /* If the tree is a leaf-data tree, and the siblings are leaves, 
      ** then there is no divider cell in apCell[]. Instead, the divider 
      ** then there is no divider cell in b.apCell[]. Instead, the divider 
      ** cell consists of the integer key for the right-most cell of 
      ** the sibling-page assembled above only.
      */
      CellInfo info;
      j--;
      pNew->xParseCell(pNew, apCell[j], &info);
      pNew->xParseCell(pNew, b.apCell[j], &info);
      pCell = pTemp;
      sz = 4 + putVarint(&pCell[4], info.nKey);
      pTemp = 0;
    }else{
      pCell -= 4;
      /* Obscure case for non-leaf-data trees: If the cell at pCell was
      ** previously stored on a leaf node, and its reported size was 4
      ** bytes, then it may actually be smaller than this 
      ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
      ** any cell). But it is important to pass the correct size to 
      ** insertCell(), so reparse the cell now.
      **
      ** Note that this can never happen in an SQLite data file, as all
      ** cells are at least 4 bytes. It only happens in b-trees used
      ** to evaluate "IN (SELECT ...)" and similar clauses.
      */
      if( szCell[j]==4 ){
      if( b.szCell[j]==4 ){
        assert(leafCorrection==4);
        sz = pParent->xCellSize(pParent, pCell);
      }
    }
    iOvflSpace += sz;
    assert( sz<=pBt->maxLocal+23 );
    assert( iOvflSpace <= (int)pBt->pageSize );

      ** only after iPg+1 has already been updated. */
      assert( cntNew[iPg]>=cntOld[iPg] || abDone[iPg+1] );

      if( iPg==0 ){
        iNew = iOld = 0;
        nNewCell = cntNew[0];
      }else{
        iOld = iPg<nOld ? (cntOld[iPg-1] + !leafData) : nCell;
        iOld = iPg<nOld ? (cntOld[iPg-1] + !leafData) : b.nCell;
        iNew = cntNew[iPg-1] + !leafData;
        nNewCell = cntNew[iPg] - iNew;
      }

      editPage(apNew[iPg], iOld, iNew, nNewCell, apCell, szCell);
      rc = editPage(apNew[iPg], iOld, iNew, nNewCell, &b);
      if( rc ) goto balance_cleanup;
      abDone[iPg]++;
      apNew[iPg]->nFree = usableSpace-szNew[iPg];
      assert( apNew[iPg]->nOverflow==0 );
      assert( apNew[iPg]->nCell==nNewCell );
    }
  }

      u32 key = get4byte(&apNew[i]->aData[8]);
      ptrmapPut(pBt, key, PTRMAP_BTREE, apNew[i]->pgno, &rc);
    }
  }

  assert( pParent->isInit );
  TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
          nOld, nNew, nCell));
          nOld, nNew, b.nCell));

  /* Free any old pages that were not reused as new pages.
  */
  for(i=nNew; i<nOld; i++){
    freePage(apOld[i], &rc);
  }

  }
#endif

  /*
  ** Cleanup before returning.
  */
balance_cleanup:
  sqlite3ScratchFree(apCell);
  sqlite3ScratchFree(b.apCell);
  for(i=0; i<nOld; i++){
    releasePage(apOld[i]);
  }
  for(i=0; i<nNew; i++){
    releasePage(apNew[i]);
  }

  if( !pPage->leaf ){
    MemPage *pLeaf = pCur->apPage[pCur->iPage];
    int nCell;
    Pgno n = pCur->apPage[iCellDepth+1]->pgno;
    unsigned char *pTmp;

    pCell = findCell(pLeaf, pLeaf->nCell-1);
    if( pCell<&pLeaf->aData[4] ) return SQLITE_CORRUPT_BKPT;
    if( NEVER(pCell<&pLeaf->aData[4]) ) return SQLITE_CORRUPT_BKPT;
    nCell = pLeaf->xCellSize(pLeaf, pCell);
    assert( MX_CELL_SIZE(pBt) >= nCell );
    pTmp = pBt->pTmpSpace;
    assert( pTmp!=0 );
    rc = sqlite3PagerWrite(pLeaf->pDbPage);
    insertCell(pPage, iCellIdx, pCell-4, nCell+4, pTmp, n, &rc);
    dropCell(pLeaf, pLeaf->nCell-1, nCell, &rc);