SQLite

** A PMA created at this point is known as a "level-0 PMA". Higher levels
** of PMAs may be created by merging existing PMAs together - for example
** merging two or more level-0 PMAs together creates a level-1 PMA.
**
** The threshold for the amount of main memory to use before flushing 
** records to a PMA is roughly the same as the limit configured for the
** page-cache of the main database. Specifically, the threshold is set to 
** the value returned by "PRAGMA main.page_size" multipled by 
** that returned by "PRAGMA main.cache_size", in bytes.
**
** If the sorter is running in single-threaded mode, then all PMAs generated
** are appended to a single temporary file. Or, if the sorter is running in
** multi-threaded mode then up to (N+1) temporary files may be opened, where
** N is the configured number of worker threads. In this case, instead of
** sorting the records and writing the PMA to a temporary file itself, the

** The MergeEngine object is used to combine two or more smaller PMAs into
** one big PMA using a merge operation.  Separate PMAs all need to be
** combined into one big PMA in order to be able to step through the sorted
** records in order.
**
** The aReadr[] array contains a PmaReader object for each of the PMAs being
** merged.  An aReadr[] object either points to a valid key or else is at EOF.
** ("EOF" means "End Of File".  When aReadr[] is at EOF there is no more data.)
** For the purposes of the paragraphs below, we assume that the array is
** actually N elements in size, where N is the smallest power of 2 greater
** to or equal to the number of PMAs being merged. The extra aReadr[] elements
** are treated as if they are empty (always at EOF).
**
** The aTree[] array is also N elements in size. The value of N is stored in
** the MergeEngine.nTree variable.

**
** In other words, each time we advance to the next sorter element, log2(N)
** key comparison operations are required, where N is the number of segments
** being merged (rounded up to the next power of 2).
*/
struct MergeEngine {
  int nTree;                 /* Used size of aTree/aReadr (power of 2) */
  SortSubtask *pTask;        /* Used by this thread only */
  int *aTree;                /* Current state of incremental merge */
  PmaReader *aReadr;         /* Array of PmaReaders to merge data from */
};

/*
** This object represents a single thread of control in a sort operation.
** Exactly VdbeSorter.nTask instances of this object are allocated
** as part of each VdbeSorter object. Instances are never allocated any
** other way. VdbeSorter.nTask is set to the number of worker threads allowed
** (see SQLITE_CONFIG_WORKER_THREADS) plus one (the main thread).  Thus for
** single-threaded operation, there is exactly one instance of this object
** and for multi-threaded operation there are two or more instances.
**
** Essentially, this structure contains all those fields of the VdbeSorter
** structure for which each thread requires a separate instance. For example,
** each thread requries its own UnpackedRecord object to unpack records in
** as part of comparison operations.
**
** Before a background thread is launched, variable bDone is set to 0. Then, 

/* Maximum number of PMAs that a single MergeEngine can merge */
#define SORTER_MAX_MERGE_COUNT 16

static int vdbeIncrSwap(IncrMerger*);
static void vdbeIncrFree(IncrMerger *);

/*
** 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->pFile, 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
** and return SQLITE_OK. Otherwise, if an error occurs, return an SQLite
** error code.
**
** The buffer returned in *ppOut is only valid until the
** next call to this function.
*/
static int vdbePmaReadBlob(
  PmaReader *p,                   /* PmaReader from which to take the blob */
  int nByte,                      /* Bytes of data to read */
  u8 **ppOut                      /* OUT: Pointer to buffer containing data */
){

** Or, if an error occurs, return an SQLite error code. The final value of
** *pp is undefined in this case.
*/
static int vdbeSorterMapFile(SortSubtask *pTask, SorterFile *pFile, u8 **pp){
  int rc = SQLITE_OK;
  if( pFile->iEof<=(i64)(pTask->pSorter->db->nMaxSorterMmap) ){
    rc = sqlite3OsFetch(pFile->pFd, 0, (int)pFile->iEof, (void**)pp);
    testcase( rc!=SQLITE_OK );
  }
  return rc;
}

/*
** Seek PmaReader pReadr to offset iOff within file pFile. Return SQLITE_OK 
** if successful, or an SQLite error code if an error occurs.
*/
static int vdbePmaReaderSeek(
  SortSubtask *pTask,             /* Task context */
  PmaReader *pReadr,              /* Reader whose cursor is to be moved */
  SorterFile *pFile,              /* Sorter file to read from */
  i64 iOff                        /* Offset in pFile */
){
  int rc = SQLITE_OK;

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

      int nRead = pgsz - iBuf;
      if( (pReadr->iReadOff + nRead) > pReadr->iEof ){
        nRead = (int)(pReadr->iEof - pReadr->iReadOff);
      }
      rc = sqlite3OsRead(
          pReadr->pFile, &pReadr->aBuffer[iBuf], nRead, pReadr->iReadOff
      );
      testcase( rc!=SQLITE_OK );
    }
  }

  return rc;
}

/*

        bEof = 0;
      }
    }

    if( bEof ){
      /* This is an EOF condition */
      vdbePmaReaderClear(pReadr);
      testcase( rc!=SQLITE_OK );
      return rc;
    }
  }

  if( rc==SQLITE_OK ){
    rc = vdbePmaReadVarint(pReadr, &nRec);
  }
  if( rc==SQLITE_OK ){
    pReadr->nKey = (int)nRec;
    rc = vdbePmaReadBlob(pReadr, (int)nRec, &pReadr->aKey);
    testcase( rc!=SQLITE_OK );
  }

  return rc;
}

/*
** Initialize PmaReader pReadr to scan through the PMA stored in file pFile

}
#else
# define vdbeSorterJoinAll(x,rcin) (rcin)
# define vdbeSorterJoinThread(pTask) SQLITE_OK
#endif

/*
** Allocate a new MergeEngine object capable of handling up to
** nReader PmaReader inputs.
**
** nReader is automatically rounded up to the next power of two.
** nReader may not exceed SORTER_MAX_MERGE_COUNT even after rounding up.
*/
static MergeEngine *vdbeMergeEngineNew(int nReader){
  int N = 2;                      /* Smallest power of two >= nReader */
  int nByte;                      /* Total bytes of space to allocate */
  MergeEngine *pNew;              /* Pointer to allocated object to return */

  assert( nReader<=SORTER_MAX_MERGE_COUNT );

  while( N<nReader ) N += N;
  nByte = sizeof(MergeEngine) + N * (sizeof(int) + sizeof(PmaReader));

  pNew = sqlite3FaultSim(100) ? 0 : (MergeEngine*)sqlite3MallocZero(nByte);
  if( pNew ){
    pNew->nTree = N;
    pNew->pTask = 0;
    pNew->aReadr = (PmaReader*)&pNew[1];
    pNew->aTree = (int*)&pNew->aReadr[N];
  }
  return pNew;
}

/*

  vdbeSorterWorkDebug(pTask, "exit");
  assert( rc!=SQLITE_OK || pList->pList==0 );
  assert( rc!=SQLITE_OK || pTask->file.iEof==iSz );
  return rc;
}

/*
** Advance the MergeEngine pMerge (passed as the second argument) to
** its next entry.  Set *pbEof to true there is no next entry because
** the MergeEngine has reached the end of all its inputs.
**
** Return SQLITE_OK if successful or an error code if an error occurs.
*/
static int vdbeMergeEngineStep(
  SortSubtask *pTask,        /* The thread in which this MergeEngine runs */
  MergeEngine *pMerger,      /* The merge engine to advance to the next row */
  int *pbEof                 /* Set TRUE at EOF.  Set false for more content */
){
  int rc;
  int iPrev = pMerger->aTree[1];/* Index of PmaReader to advance */

  /* A MergeEngine object is only used by a single thread */
  assert( pMerger->pTask==0 || pMerger->pTask==pTask );
  pMerger->pTask = pTask;

  /* Advance the current PmaReader */
  rc = vdbePmaReaderNext(&pMerger->aReadr[iPrev]);

  /* Update contents of aTree[] */
  if( rc==SQLITE_OK ){
    int i;                      /* Index of aTree[] to recalculate */

    ** In either case exit the loop. */
    if( pReader->pFile==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);
  }

  rc2 = vdbePmaWriterFinish(&writer, &pOut->iEof);
  if( rc==SQLITE_OK ) rc = rc2;
  vdbeSorterPopulateDebug(pTask, "exit");
  return rc;
}

** can be used to incrementally merge all PMAs on disk.
**
** If successful, SQLITE_OK is returned and *ppOut set to point to the
** MergeEngine object at the root of the tree before returning. Or, if an
** error occurs, an SQLite error code is returned and the final value 
** of *ppOut is undefined.
*/
static int vdbeSorterMergeTreeBuild(
  VdbeSorter *pSorter,       /* The VDBE cursor that implements the sort */
  MergeEngine **ppOut        /* Write the MergeEngine here */
){
  MergeEngine *pMain = 0;
  int rc = SQLITE_OK;
  int iTask;

#if SQLITE_MAX_WORKER_THREADS>0
  /* If the sorter uses more than one task, then create the top-level 
  ** MergeEngine here. This MergeEngine will read data from exactly 

#if SQLITE_MAX_WORKER_THREADS>0
    if( pSorter->bUseThreads ){
      rc = vdbePmaReaderNext(pSorter->pReader);
      *pbEof = (pSorter->pReader->pFile==0);
    }else
#endif
    /*if( !pSorter->bUseThreads )*/ {
      rc = vdbeMergeEngineStep(&pSorter->aTask[0], pSorter->pMerger, pbEof);
    }
  }else{
    SorterRecord *pFree = pSorter->list.pList;
    pSorter->list.pList = pFree->u.pNext;
    pFree->u.pNext = 0;
    if( pSorter->list.aMemory==0 ) vdbeSorterRecordFree(db, pFree);
    *pbEof = !pSorter->list.pList;