SQLite

  u8 nTask;                       /* Size of aTask[] array */
  SortSubtask aTask[1];           /* One or more subtasks */
};

/*
** An instance of the following object is used to read records out of a
** PMA, in sorted order.  The next key to be read is cached in nKey/aKey.
** aKey might point into aMap or into aBuffer.  If neither of those locations
** contain a contiguous representation of the key, then aAlloc is allocated
** and the key is copied into aAlloc and aKey is made to poitn to aAlloc.
**
** pFd==0 at EOF.
*/
struct PmaReader {
  i64 iReadOff;               /* Current read offset */
  i64 iEof;                   /* 1 byte past EOF for this PmaReader */
  int nAlloc;                 /* Bytes of space at aAlloc */
  int nKey;                   /* Number of bytes in key */
  sqlite3_file *pFd;          /* File handle we are reading from */
  u8 *aAlloc;                 /* Space for aKey if aBuffer and pMap wont work */
  u8 *aKey;                   /* Pointer to current key */
  u8 *aBuffer;                /* Current read buffer */
  int nBuffer;                /* Size of read buffer in bytes */
  u8 *aMap;                   /* Pointer to mapping of entire file */
  IncrMerger *pIncr;          /* Incremental merger */
};

/*
** Normally, a PmaReader object iterates through an existing PMA stored 
** within a temp file. However, if the PmaReader.pIncr variable points to
** an object of the following type, it may be used to iterate/merge through
** multiple PMAs simultaneously.

struct PmaWriter {
  int eFWErr;                     /* Non-zero if in an error state */
  u8 *aBuffer;                    /* Pointer to write buffer */
  int nBuffer;                    /* Size of write buffer in bytes */
  int iBufStart;                  /* First byte of buffer to write */
  int iBufEnd;                    /* Last byte of buffer to write */
  i64 iWriteOff;                  /* Offset of start of buffer in file */
  sqlite3_file *pFd;              /* File handle to write to */
};

/*
** This object is the header on a single record while that record is being
** held in memory and prior to being written out as part of a PMA.
**
** How the linked list is connected depends on how memory is being managed

/*
** Free all memory belonging to the PmaReader object passed as the
** argument. All structure fields are set to zero before returning.
*/
static void vdbePmaReaderClear(PmaReader *pReadr){
  sqlite3_free(pReadr->aAlloc);
  sqlite3_free(pReadr->aBuffer);
  if( pReadr->aMap ) sqlite3OsUnfetch(pReadr->pFd, 0, pReadr->aMap);
  vdbeIncrFree(pReadr->pIncr);
  memset(pReadr, 0, sizeof(PmaReader));
}

/*
** Read the next nByte bytes of data from the PMA p.
** If successful, set *ppOut to point to a buffer containing the data

      nRead = p->nBuffer;
    }else{
      nRead = (int)(p->iEof - p->iReadOff);
    }
    assert( nRead>0 );

    /* Readr data from the file. Return early if an error occurs. */
    rc = sqlite3OsRead(p->pFd, p->aBuffer, nRead, p->iReadOff);
    assert( rc!=SQLITE_IOERR_SHORT_READ );
    if( rc!=SQLITE_OK ) return rc;
  }
  nAvail = p->nBuffer - iBuf; 

  if( nByte<=nAvail ){
    /* The requested data is available in the in-memory buffer. In this

  i64 iOff                        /* Offset in pFile */
){
  int rc = SQLITE_OK;

  assert( pReadr->pIncr==0 || pReadr->pIncr->bEof==0 );

  if( pReadr->aMap ){
    sqlite3OsUnfetch(pReadr->pFd, 0, pReadr->aMap);
    pReadr->aMap = 0;
  }
  pReadr->iReadOff = iOff;
  pReadr->iEof = pFile->iEof;
  pReadr->pFd = pFile->pFd;

  rc = vdbeSorterMapFile(pTask, pFile, &pReadr->aMap);
  if( rc==SQLITE_OK && pReadr->aMap==0 ){
    int pgsz = pTask->pSorter->pgsz;
    int iBuf = pReadr->iReadOff % pgsz;
    if( pReadr->aBuffer==0 ){
      pReadr->aBuffer = (u8*)sqlite3Malloc(pgsz);
      if( pReadr->aBuffer==0 ) rc = SQLITE_NOMEM;
      pReadr->nBuffer = pgsz;
    }
    if( rc==SQLITE_OK && iBuf ){
      int nRead = pgsz - iBuf;
      if( (pReadr->iReadOff + nRead) > pReadr->iEof ){
        nRead = (int)(pReadr->iEof - pReadr->iReadOff);
      }
      rc = sqlite3OsRead(
          pReadr->pFd, &pReadr->aBuffer[iBuf], nRead, pReadr->iReadOff
      );
      testcase( rc!=SQLITE_OK );
    }
  }

  return rc;
}

    i1 = pMerger->aTree[iOut*2];
    i2 = pMerger->aTree[iOut*2+1];
  }

  p1 = &pMerger->aReadr[i1];
  p2 = &pMerger->aReadr[i2];

  if( p1->pFd==0 ){
    iRes = i2;
  }else if( p2->pFd==0 ){
    iRes = i1;
  }else{
    int res;
    assert( pTask->pUnpacked!=0 );  /* allocated in vdbeSortSubtaskMain() */
    res = vdbeSorterCompare(
        pTask, p1->aKey, p1->nKey, p2->aKey, p2->nKey
    );

** is guaranteed to be nByte bytes or smaller in size. This function
** attempts to extend the file to nByte bytes in size and to ensure that
** the VFS has memory mapped it.
**
** Whether or not the file does end up memory mapped of course depends on
** the specific VFS implementation.
*/
static void vdbeSorterExtendFile(sqlite3 *db, sqlite3_file *pFd, i64 nByte){
  if( nByte<=(i64)(db->nMaxSorterMmap) ){
    int rc = sqlite3OsTruncate(pFd, nByte);
    if( rc==SQLITE_OK ){
      void *p = 0;
      sqlite3OsFetch(pFd, 0, (int)nByte, &p);
      sqlite3OsUnfetch(pFd, 0, p);
    }
  }
}
#else
# define vdbeSorterExtendFile(x,y,z) SQLITE_OK
#endif

/*
** Allocate space for a file-handle and open a temporary file. If successful,
** set *ppFd to point to the malloc'd file-handle and return SQLITE_OK.
** Otherwise, set *ppFd to 0 and return an SQLite error code.
*/
static int vdbeSorterOpenTempFile(
  sqlite3 *db,                    /* Database handle doing sort */
  i64 nExtend,                    /* Attempt to extend file to this size */
  sqlite3_file **ppFd
){
  int rc;
  rc = sqlite3OsOpenMalloc(db->pVfs, 0, ppFd,
      SQLITE_OPEN_TEMP_JOURNAL |
      SQLITE_OPEN_READWRITE    | SQLITE_OPEN_CREATE |
      SQLITE_OPEN_EXCLUSIVE    | SQLITE_OPEN_DELETEONCLOSE, &rc
  );
  if( rc==SQLITE_OK ){
    i64 max = SQLITE_MAX_MMAP_SIZE;
    sqlite3OsFileControlHint(*ppFd, SQLITE_FCNTL_MMAP_SIZE, (void*)&max);
    if( nExtend>0 ){
      vdbeSorterExtendFile(db, *ppFd, nExtend);
    }
  }
  return rc;
}

/*
** If it has not already been allocated, allocate the UnpackedRecord 

  return pTask->pUnpacked->errCode;
}

/*
** Initialize a PMA-writer object.
*/
static void vdbePmaWriterInit(
  sqlite3_file *pFd,              /* File handle to write to */
  PmaWriter *p,                   /* Object to populate */
  int nBuf,                       /* Buffer size */
  i64 iStart                      /* Offset of pFd to begin writing at */
){
  memset(p, 0, sizeof(PmaWriter));
  p->aBuffer = (u8*)sqlite3Malloc(nBuf);
  if( !p->aBuffer ){
    p->eFWErr = SQLITE_NOMEM;
  }else{
    p->iBufEnd = p->iBufStart = (iStart % nBuf);
    p->iWriteOff = iStart - p->iBufStart;
    p->nBuffer = nBuf;
    p->pFd = pFd;
  }
}

/*
** Write nData bytes of data to the PMA. Return SQLITE_OK
** if successful, or an SQLite error code if an error occurs.
*/
static void vdbePmaWriteBlob(PmaWriter *p, u8 *pData, int nData){
  int nRem = nData;
  while( nRem>0 && p->eFWErr==0 ){
    int nCopy = nRem;
    if( nCopy>(p->nBuffer - p->iBufEnd) ){
      nCopy = p->nBuffer - p->iBufEnd;
    }

    memcpy(&p->aBuffer[p->iBufEnd], &pData[nData-nRem], nCopy);
    p->iBufEnd += nCopy;
    if( p->iBufEnd==p->nBuffer ){
      p->eFWErr = sqlite3OsWrite(p->pFd, 
          &p->aBuffer[p->iBufStart], p->iBufEnd - p->iBufStart, 
          p->iWriteOff + p->iBufStart
      );
      p->iBufStart = p->iBufEnd = 0;
      p->iWriteOff += p->nBuffer;
    }
    assert( p->iBufEnd<p->nBuffer );

**
** Before returning, set *piEof to the offset immediately following the
** last byte written to the file.
*/
static int vdbePmaWriterFinish(PmaWriter *p, i64 *piEof){
  int rc;
  if( p->eFWErr==0 && ALWAYS(p->aBuffer) && p->iBufEnd>p->iBufStart ){
    p->eFWErr = sqlite3OsWrite(p->pFd, 
        &p->aBuffer[p->iBufStart], p->iBufEnd - p->iBufStart, 
        p->iWriteOff + p->iBufStart
    );
  }
  *piEof = (p->iWriteOff + p->iBufEnd);
  sqlite3_free(p->aBuffer);
  rc = p->eFWErr;

    pReadr1 = &pMerger->aReadr[(iPrev & 0xFFFE)];
    pReadr2 = &pMerger->aReadr[(iPrev | 0x0001)];
    pKey2 = pReadr2->aKey;

    for(i=(pMerger->nTree+iPrev)/2; i>0; i=i/2){
      /* Compare pReadr1 and pReadr2. Store the result in variable iRes. */
      int iRes;
      if( pReadr1->pFd==0 ){
        iRes = +1;
      }else if( pReadr2->pFd==0 ){
        iRes = -1;
      }else{
        iRes = vdbeSorterCompare(pTask, 
            pReadr1->aKey, pReadr1->nKey, pKey2, pReadr2->nKey
        );
      }

      ** is sorted from oldest to newest, so pReadr1 contains older values
      ** than pReadr2 iff (pReadr1<pReadr2).  */
      if( iRes<0 || (iRes==0 && pReadr1<pReadr2) ){
        pMerger->aTree[i] = (int)(pReadr1 - pMerger->aReadr);
        pReadr2 = &pMerger->aReadr[ pMerger->aTree[i ^ 0x0001] ];
        pKey2 = pReadr2->aKey;
      }else{
        if( pReadr1->pFd ) pKey2 = 0;
        pMerger->aTree[i] = (int)(pReadr2 - pMerger->aReadr);
        pReadr1 = &pMerger->aReadr[ pMerger->aTree[i ^ 0x0001] ];
      }
    }
    *pbEof = (pMerger->aReadr[pMerger->aTree[1]].pFd==0);
  }

  return (rc==SQLITE_OK ? pTask->pUnpacked->errCode : rc);
}

#if SQLITE_MAX_WORKER_THREADS>0
/*

    int dummy;
    PmaReader *pReader = &pMerger->aReadr[ pMerger->aTree[1] ];
    int nKey = pReader->nKey;
    i64 iEof = writer.iWriteOff + writer.iBufEnd;

    /* Check if the output file is full or if the input has been exhausted.
    ** In either case exit the loop. */
    if( pReader->pFd==0 ) break;
    if( (iEof + nKey + sqlite3VarintLen(nKey))>(iStart + pIncr->mxSz) ) break;

    /* Write the next key to the output. */
    vdbePmaWriteVarint(&writer, nKey);
    vdbePmaWriteBlob(&writer, pReader->aKey, nKey);
    rc = vdbeMergeEngineStep(pTask, pIncr->pMerger, &dummy);
  }

/*
** Allocate and return a new IncrMerger object to read data from pMerger.
**
** If an OOM condition is encountered, return NULL. In this case free the
** pMerger argument before returning.
*/
static int vdbeIncrNew(
  SortSubtask *pTask,     /* The thread that will be using the new IncrMerger */
  MergeEngine *pMerger,   /* The MergeEngine that the IncrMerger will control */
  IncrMerger **ppOut      /* Write the new IncrMerger here */
){
  int rc = SQLITE_OK;
  IncrMerger *pIncr = *ppOut = (IncrMerger*)
       (sqlite3FaultSim(100) ? 0 : sqlite3MallocZero(sizeof(*pIncr)));
  if( pIncr ){
    pIncr->pMerger = pMerger;
    pIncr->pTask = pTask;

#define INCRINIT_NORMAL 0
#define INCRINIT_TASK   1
#define INCRINIT_ROOT   2
static int vdbePmaReaderIncrInit(PmaReader *pReadr, int eMode);

/*
** Initialize the MergeEngine object passed as the second argument. Once this
** function returns, the first key of merged data may be read from the 
** MergeEngine object in the usual fashion.
**
** If argument eMode is INCRINIT_ROOT, then it is assumed that any IncrMerge
** objects attached to the PmaReader objects that the merger reads from have
** already been populated, but that they have not yet populated aFile[0] and
** set the PmaReader objects up to read from it. In this case all that is
** required is to call vdbePmaReaderNext() on each PmaReader to point it at
** its first key.
**
** Otherwise, if eMode is any value other than INCRINIT_ROOT, then use 
** vdbePmaReaderIncrInit() to initialize each PmaReader that feeds data 
** to pMerger.
**
** SQLITE_OK is returned if successful, or an SQLite error code otherwise.
*/
static int vdbeIncrInitMerger(
  SortSubtask *pTask,             /* Thread that will run pMerger */
  MergeEngine *pMerger,           /* MergeEngine to initialize */
  int eMode                       /* One of the INCRINIT_XXX constants */
){
  int rc = SQLITE_OK;             /* Return code */
  int i;                          /* For looping over PmaReader objects */
  int nTree = pMerger->nTree;

  /* Verify that the MergeEngine is assigned to a single thread */
  assert( pMerger->pTask==0 || pMerger->pTask==pTask );
  pMerger->pTask = pTask;

  for(i=0; rc==SQLITE_OK && i<nTree; i++){
    if( eMode==INCRINIT_ROOT ){
      /* PmaReaders should be normally initialized in order, as if they are
      ** reading from the same temp file this makes for more linear file IO.
      ** However, in the INCRINIT_ROOT case, if PmaReader aReadr[nTask-1] is
      ** in use it will block the vdbePmaReaderNext() call while it uses

/*
** If the PmaReader passed as the first argument is not an incremental-reader
** (if pReadr->pIncr==0), then this function is a no-op. Otherwise, it serves
** to open and/or initialize the temp file related fields of the IncrMerge
** object at (pReadr->pIncr).
**
** If argument eMode is set to INCRINIT_NORMAL, then all PmaReaders
** in the sub-tree headed by pReadr are also initialized. Data is then loaded
** into the buffers belonging to pReadr and it is set to
** point to the first key in its range.
**
** If argument eMode is set to INCRINIT_TASK, then pReadr is guaranteed
** to be a multi-threaded PmaReader and this function is being called in a
** background thread. In this case all PmaReaders in the sub-tree are 
** initialized as for INCRINIT_NORMAL and the aFile[1] buffer belonging to
** pReadr is populated. However, pReadr itself is not set up to point
** to its first key. A call to vdbePmaReaderNext() is still required to do
** that. 
**
** The reason this function does not call vdbePmaReaderNext() immediately 
** in the INCRINIT_TASK case is that vdbePmaReaderNext() assumes that it has
** to block on thread (pTask->thread) before accessing aFile[1]. But, since
** this entire function is being run by thread (pTask->thread), that will
** lead to the current background thread attempting to join itself.
**
** Finally, if argument eMode is set to INCRINIT_ROOT, it may be assumed
** that pReadr->pIncr is a multi-threaded IncrMerge objects, and that all
** child-trees have already been initialized using IncrInit(INCRINIT_TASK).

  if( pSorter->bUsePMA ){
    assert( pSorter->pReader==0 || pSorter->pMerger==0 );
    assert( pSorter->bUseThreads==0 || pSorter->pReader );
    assert( pSorter->bUseThreads==1 || pSorter->pMerger );
#if SQLITE_MAX_WORKER_THREADS>0
    if( pSorter->bUseThreads ){
      rc = vdbePmaReaderNext(pSorter->pReader);
      *pbEof = (pSorter->pReader->pFd==0);
    }else
#endif
    /*if( !pSorter->bUseThreads )*/ {
      rc = vdbeMergeEngineStep(&pSorter->aTask[0], pSorter->pMerger, pbEof);
    }
  }else{
    SorterRecord *pFree = pSorter->list.pList;