/ Check-in [79326d6e]
Login

Many hyperlinks are disabled.
Use anonymous login to enable hyperlinks.

Overview
Comment:Optimizations for VACUUM, CREATE INDEX and some cases of ORDER BY.
Downloads: Tarball | ZIP archive | SQL archive
Timelines: family | ancestors | descendants | both | trunk
Files: files | file ages | folders
SHA1: 79326d6eece926fd1c148b29f0b726208d8b44c0
User & Date: dan 2015-04-11 20:44:28
References
2015-04-12
21:28
Update the version number to 3.8.10 because of the new sort optimizations from check-in [79326d6eec]. check-in: bd2cfa1d user: drh tags: trunk
Context
2015-04-12
01:22
Avoid problems when a query has a GROUP BY and an ORDER BY but no FROM clause. check-in: e527d96a user: drh tags: trunk
2015-04-11
20:44
Optimizations for VACUUM, CREATE INDEX and some cases of ORDER BY. check-in: 79326d6e user: dan tags: trunk
20:20
Fix a problem with sorting large amounts of partially ordered data. check-in: acca97ef user: dan tags: trunk
18:45
Merge latest trunk changes with this branch. Closed-Leaf check-in: cf7590f6 user: dan tags: sorter-opt
Changes
Hide Diffs Unified Diffs Ignore Whitespace Patch

Changes to src/build.c.

2759
2760
2761
2762
2763
2764
2765

2766
2767
2768
2769
2770
2771
2772
2773
    sqlite3VdbeAddOp4Int(v, OP_SorterCompare, iSorter, j2, regRecord,
                         pIndex->nKeyCol); VdbeCoverage(v);
    sqlite3UniqueConstraint(pParse, OE_Abort, pIndex);
  }else{
    addr2 = sqlite3VdbeCurrentAddr(v);
  }
  sqlite3VdbeAddOp3(v, OP_SorterData, iSorter, regRecord, iIdx);

  sqlite3VdbeAddOp3(v, OP_IdxInsert, iIdx, regRecord, 1);
  sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT);
  sqlite3ReleaseTempReg(pParse, regRecord);
  sqlite3VdbeAddOp2(v, OP_SorterNext, iSorter, addr2); VdbeCoverage(v);
  sqlite3VdbeJumpHere(v, addr1);

  sqlite3VdbeAddOp1(v, OP_Close, iTab);
  sqlite3VdbeAddOp1(v, OP_Close, iIdx);







>
|







2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
    sqlite3VdbeAddOp4Int(v, OP_SorterCompare, iSorter, j2, regRecord,
                         pIndex->nKeyCol); VdbeCoverage(v);
    sqlite3UniqueConstraint(pParse, OE_Abort, pIndex);
  }else{
    addr2 = sqlite3VdbeCurrentAddr(v);
  }
  sqlite3VdbeAddOp3(v, OP_SorterData, iSorter, regRecord, iIdx);
  sqlite3VdbeAddOp3(v, OP_Last, iIdx, 0, -1);
  sqlite3VdbeAddOp3(v, OP_IdxInsert, iIdx, regRecord, 0);
  sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT);
  sqlite3ReleaseTempReg(pParse, regRecord);
  sqlite3VdbeAddOp2(v, OP_SorterNext, iSorter, addr2); VdbeCoverage(v);
  sqlite3VdbeJumpHere(v, addr1);

  sqlite3VdbeAddOp1(v, OP_Close, iTab);
  sqlite3VdbeAddOp1(v, OP_Close, iIdx);

Changes to src/insert.c.

1761
1762
1763
1764
1765
1766
1767

1768
1769
1770
1771
1772
1773
1774
....
1908
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939

1940
1941
1942
1943
1944
1945




1946
1947
1948
1949
1950
1951
1952
1953
1954
....
1983
1984
1985
1986
1987
1988
1989

1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002

























2003

2004
2005
2006
2007
2008
2009
2010
static int xferOptimization(
  Parse *pParse,        /* Parser context */
  Table *pDest,         /* The table we are inserting into */
  Select *pSelect,      /* A SELECT statement to use as the data source */
  int onError,          /* How to handle constraint errors */
  int iDbDest           /* The database of pDest */
){

  ExprList *pEList;                /* The result set of the SELECT */
  Table *pSrc;                     /* The table in the FROM clause of SELECT */
  Index *pSrcIdx, *pDestIdx;       /* Source and destination indices */
  struct SrcList_item *pItem;      /* An element of pSelect->pSrc */
  int i;                           /* Loop counter */
  int iDbSrc;                      /* The database of pSrc */
  int iSrc, iDest;                 /* Cursors from source and destination */
................................................................................
  /* Disallow the transfer optimization if the destination table constains
  ** any foreign key constraints.  This is more restrictive than necessary.
  ** But the main beneficiary of the transfer optimization is the VACUUM 
  ** command, and the VACUUM command disables foreign key constraints.  So
  ** the extra complication to make this rule less restrictive is probably
  ** not worth the effort.  Ticket [6284df89debdfa61db8073e062908af0c9b6118e]
  */
  if( (pParse->db->flags & SQLITE_ForeignKeys)!=0 && pDest->pFKey!=0 ){
    return 0;
  }
#endif
  if( (pParse->db->flags & SQLITE_CountRows)!=0 ){
    return 0;  /* xfer opt does not play well with PRAGMA count_changes */
  }

  /* If we get this far, it means that the xfer optimization is at
  ** least a possibility, though it might only work if the destination
  ** table (tab1) is initially empty.
  */
#ifdef SQLITE_TEST
  sqlite3_xferopt_count++;
#endif
  iDbSrc = sqlite3SchemaToIndex(pParse->db, pSrc->pSchema);
  v = sqlite3GetVdbe(pParse);
  sqlite3CodeVerifySchema(pParse, iDbSrc);
  iSrc = pParse->nTab++;
  iDest = pParse->nTab++;
  regAutoinc = autoIncBegin(pParse, iDbDest, pDest);
  regData = sqlite3GetTempReg(pParse);
  regRowid = sqlite3GetTempReg(pParse);
  sqlite3OpenTable(pParse, iDest, iDbDest, pDest, OP_OpenWrite);
  assert( HasRowid(pDest) || destHasUniqueIdx );

  if( (pDest->iPKey<0 && pDest->pIndex!=0)          /* (1) */
   || destHasUniqueIdx                              /* (2) */
   || (onError!=OE_Abort && onError!=OE_Rollback)   /* (3) */
  ){
    /* In some circumstances, we are able to run the xfer optimization
    ** only if the destination table is initially empty.  This code makes




    ** that determination.  Conditions under which the destination must
    ** be empty:
    **
    ** (1) There is no INTEGER PRIMARY KEY but there are indices.
    **     (If the destination is not initially empty, the rowid fields
    **     of index entries might need to change.)
    **
    ** (2) The destination has a unique index.  (The xfer optimization 
    **     is unable to test uniqueness.)
................................................................................
    sqlite3VdbeAddOp2(v, OP_Close, iSrc, 0);
    sqlite3VdbeAddOp2(v, OP_Close, iDest, 0);
  }else{
    sqlite3TableLock(pParse, iDbDest, pDest->tnum, 1, pDest->zName);
    sqlite3TableLock(pParse, iDbSrc, pSrc->tnum, 0, pSrc->zName);
  }
  for(pDestIdx=pDest->pIndex; pDestIdx; pDestIdx=pDestIdx->pNext){

    for(pSrcIdx=pSrc->pIndex; ALWAYS(pSrcIdx); pSrcIdx=pSrcIdx->pNext){
      if( xferCompatibleIndex(pDestIdx, pSrcIdx) ) break;
    }
    assert( pSrcIdx );
    sqlite3VdbeAddOp3(v, OP_OpenRead, iSrc, pSrcIdx->tnum, iDbSrc);
    sqlite3VdbeSetP4KeyInfo(pParse, pSrcIdx);
    VdbeComment((v, "%s", pSrcIdx->zName));
    sqlite3VdbeAddOp3(v, OP_OpenWrite, iDest, pDestIdx->tnum, iDbDest);
    sqlite3VdbeSetP4KeyInfo(pParse, pDestIdx);
    sqlite3VdbeChangeP5(v, OPFLAG_BULKCSR);
    VdbeComment((v, "%s", pDestIdx->zName));
    addr1 = sqlite3VdbeAddOp2(v, OP_Rewind, iSrc, 0); VdbeCoverage(v);
    sqlite3VdbeAddOp2(v, OP_RowKey, iSrc, regData);

























    sqlite3VdbeAddOp3(v, OP_IdxInsert, iDest, regData, 1);

    sqlite3VdbeAddOp2(v, OP_Next, iSrc, addr1+1); VdbeCoverage(v);
    sqlite3VdbeJumpHere(v, addr1);
    sqlite3VdbeAddOp2(v, OP_Close, iSrc, 0);
    sqlite3VdbeAddOp2(v, OP_Close, iDest, 0);
  }
  if( emptySrcTest ) sqlite3VdbeJumpHere(v, emptySrcTest);
  sqlite3ReleaseTempReg(pParse, regRowid);







>







 







|



|










|









>
|


|

|
>
>
>
>
|
<







 







>













>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>

>







1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
....
1909
1910
1911
1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952

1953
1954
1955
1956
1957
1958
1959
....
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
static int xferOptimization(
  Parse *pParse,        /* Parser context */
  Table *pDest,         /* The table we are inserting into */
  Select *pSelect,      /* A SELECT statement to use as the data source */
  int onError,          /* How to handle constraint errors */
  int iDbDest           /* The database of pDest */
){
  sqlite3 *db = pParse->db;
  ExprList *pEList;                /* The result set of the SELECT */
  Table *pSrc;                     /* The table in the FROM clause of SELECT */
  Index *pSrcIdx, *pDestIdx;       /* Source and destination indices */
  struct SrcList_item *pItem;      /* An element of pSelect->pSrc */
  int i;                           /* Loop counter */
  int iDbSrc;                      /* The database of pSrc */
  int iSrc, iDest;                 /* Cursors from source and destination */
................................................................................
  /* Disallow the transfer optimization if the destination table constains
  ** any foreign key constraints.  This is more restrictive than necessary.
  ** But the main beneficiary of the transfer optimization is the VACUUM 
  ** command, and the VACUUM command disables foreign key constraints.  So
  ** the extra complication to make this rule less restrictive is probably
  ** not worth the effort.  Ticket [6284df89debdfa61db8073e062908af0c9b6118e]
  */
  if( (db->flags & SQLITE_ForeignKeys)!=0 && pDest->pFKey!=0 ){
    return 0;
  }
#endif
  if( (db->flags & SQLITE_CountRows)!=0 ){
    return 0;  /* xfer opt does not play well with PRAGMA count_changes */
  }

  /* If we get this far, it means that the xfer optimization is at
  ** least a possibility, though it might only work if the destination
  ** table (tab1) is initially empty.
  */
#ifdef SQLITE_TEST
  sqlite3_xferopt_count++;
#endif
  iDbSrc = sqlite3SchemaToIndex(db, pSrc->pSchema);
  v = sqlite3GetVdbe(pParse);
  sqlite3CodeVerifySchema(pParse, iDbSrc);
  iSrc = pParse->nTab++;
  iDest = pParse->nTab++;
  regAutoinc = autoIncBegin(pParse, iDbDest, pDest);
  regData = sqlite3GetTempReg(pParse);
  regRowid = sqlite3GetTempReg(pParse);
  sqlite3OpenTable(pParse, iDest, iDbDest, pDest, OP_OpenWrite);
  assert( HasRowid(pDest) || destHasUniqueIdx );
  if( (db->flags & SQLITE_Vacuum)==0 && (
      (pDest->iPKey<0 && pDest->pIndex!=0)          /* (1) */
   || destHasUniqueIdx                              /* (2) */
   || (onError!=OE_Abort && onError!=OE_Rollback)   /* (3) */
  )){
    /* In some circumstances, we are able to run the xfer optimization
    ** only if the destination table is initially empty. Unless the
    ** SQLITE_Vacuum flag is set, this block generates code to make
    ** that determination. If SQLITE_Vacuum is set, then the destination
    ** table is always empty.
    **
    ** Conditions under which the destination must be empty:

    **
    ** (1) There is no INTEGER PRIMARY KEY but there are indices.
    **     (If the destination is not initially empty, the rowid fields
    **     of index entries might need to change.)
    **
    ** (2) The destination has a unique index.  (The xfer optimization 
    **     is unable to test uniqueness.)
................................................................................
    sqlite3VdbeAddOp2(v, OP_Close, iSrc, 0);
    sqlite3VdbeAddOp2(v, OP_Close, iDest, 0);
  }else{
    sqlite3TableLock(pParse, iDbDest, pDest->tnum, 1, pDest->zName);
    sqlite3TableLock(pParse, iDbSrc, pSrc->tnum, 0, pSrc->zName);
  }
  for(pDestIdx=pDest->pIndex; pDestIdx; pDestIdx=pDestIdx->pNext){
    u8 useSeekResult = 0;
    for(pSrcIdx=pSrc->pIndex; ALWAYS(pSrcIdx); pSrcIdx=pSrcIdx->pNext){
      if( xferCompatibleIndex(pDestIdx, pSrcIdx) ) break;
    }
    assert( pSrcIdx );
    sqlite3VdbeAddOp3(v, OP_OpenRead, iSrc, pSrcIdx->tnum, iDbSrc);
    sqlite3VdbeSetP4KeyInfo(pParse, pSrcIdx);
    VdbeComment((v, "%s", pSrcIdx->zName));
    sqlite3VdbeAddOp3(v, OP_OpenWrite, iDest, pDestIdx->tnum, iDbDest);
    sqlite3VdbeSetP4KeyInfo(pParse, pDestIdx);
    sqlite3VdbeChangeP5(v, OPFLAG_BULKCSR);
    VdbeComment((v, "%s", pDestIdx->zName));
    addr1 = sqlite3VdbeAddOp2(v, OP_Rewind, iSrc, 0); VdbeCoverage(v);
    sqlite3VdbeAddOp2(v, OP_RowKey, iSrc, regData);
    if( db->flags & SQLITE_Vacuum ){
      /* This INSERT command is part of a VACUUM operation, which guarantees
      ** that the destination table is empty. If all indexed columns use
      ** collation sequence BINARY, then it can also be assumed that the
      ** index will be populated by inserting keys in strictly sorted 
      ** order. In this case, instead of seeking within the b-tree as part
      ** of every OP_IdxInsert opcode, an OP_Last is added before the
      ** OP_IdxInsert to seek to the point within the b-tree where each key 
      ** should be inserted. This is faster.
      **
      ** If any of the indexed columns use a collation sequence other than
      ** BINARY, this optimization is disabled. This is because the user 
      ** might change the definition of a collation sequence and then run
      ** a VACUUM command. In that case keys may not be written in strictly
      ** sorted order.  */
      int i;
      for(i=0; i<pSrcIdx->nColumn; i++){
        char *zColl = pSrcIdx->azColl[i];
        if( zColl && sqlite3_stricmp("BINARY", zColl) ) break;
      }
      if( i==pSrcIdx->nColumn ){
        useSeekResult = OPFLAG_USESEEKRESULT;
        sqlite3VdbeAddOp3(v, OP_Last, iDest, 0, -1);
      }
    }
    sqlite3VdbeAddOp3(v, OP_IdxInsert, iDest, regData, 1);
    sqlite3VdbeChangeP5(v, useSeekResult);
    sqlite3VdbeAddOp2(v, OP_Next, iSrc, addr1+1); VdbeCoverage(v);
    sqlite3VdbeJumpHere(v, addr1);
    sqlite3VdbeAddOp2(v, OP_Close, iSrc, 0);
    sqlite3VdbeAddOp2(v, OP_Close, iDest, 0);
  }
  if( emptySrcTest ) sqlite3VdbeJumpHere(v, emptySrcTest);
  sqlite3ReleaseTempReg(pParse, regRowid);

Changes to src/sqliteInt.h.

1222
1223
1224
1225
1226
1227
1228

1229
1230
1231
1232
1233
1234
1235
#define SQLITE_AutoIndex      0x00100000  /* Enable automatic indexes */
#define SQLITE_PreferBuiltin  0x00200000  /* Preference to built-in funcs */
#define SQLITE_LoadExtension  0x00400000  /* Enable load_extension */
#define SQLITE_EnableTrigger  0x00800000  /* True to enable triggers */
#define SQLITE_DeferFKs       0x01000000  /* Defer all FK constraints */
#define SQLITE_QueryOnly      0x02000000  /* Disable database changes */
#define SQLITE_VdbeEQP        0x04000000  /* Debug EXPLAIN QUERY PLAN */



/*
** Bits of the sqlite3.dbOptFlags field that are used by the
** sqlite3_test_control(SQLITE_TESTCTRL_OPTIMIZATIONS,...) interface to
** selectively disable various optimizations.
*/







>







1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
#define SQLITE_AutoIndex      0x00100000  /* Enable automatic indexes */
#define SQLITE_PreferBuiltin  0x00200000  /* Preference to built-in funcs */
#define SQLITE_LoadExtension  0x00400000  /* Enable load_extension */
#define SQLITE_EnableTrigger  0x00800000  /* True to enable triggers */
#define SQLITE_DeferFKs       0x01000000  /* Defer all FK constraints */
#define SQLITE_QueryOnly      0x02000000  /* Disable database changes */
#define SQLITE_VdbeEQP        0x04000000  /* Debug EXPLAIN QUERY PLAN */
#define SQLITE_Vacuum         0x08000000  /* Currently in a VACUUM */


/*
** Bits of the sqlite3.dbOptFlags field that are used by the
** sqlite3_test_control(SQLITE_TESTCTRL_OPTIMIZATIONS,...) interface to
** selectively disable various optimizations.
*/

Changes to src/vacuum.c.

246
247
248
249
250
251
252


253
254
255
256
257
258
259


260
261
262
263
264
265
266
      "  FROM sqlite_master WHERE sql LIKE 'CREATE UNIQUE INDEX %'");
  if( rc!=SQLITE_OK ) goto end_of_vacuum;

  /* Loop through the tables in the main database. For each, do
  ** an "INSERT INTO vacuum_db.xxx SELECT * FROM main.xxx;" to copy
  ** the contents to the temporary database.
  */


  rc = execExecSql(db, pzErrMsg,
      "SELECT 'INSERT INTO vacuum_db.' || quote(name) "
      "|| ' SELECT * FROM main.' || quote(name) || ';'"
      "FROM main.sqlite_master "
      "WHERE type = 'table' AND name!='sqlite_sequence' "
      "  AND coalesce(rootpage,1)>0"
  );


  if( rc!=SQLITE_OK ) goto end_of_vacuum;

  /* Copy over the sequence table
  */
  rc = execExecSql(db, pzErrMsg,
      "SELECT 'DELETE FROM vacuum_db.' || quote(name) || ';' "
      "FROM vacuum_db.sqlite_master WHERE name='sqlite_sequence' "







>
>







>
>







246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
      "  FROM sqlite_master WHERE sql LIKE 'CREATE UNIQUE INDEX %'");
  if( rc!=SQLITE_OK ) goto end_of_vacuum;

  /* Loop through the tables in the main database. For each, do
  ** an "INSERT INTO vacuum_db.xxx SELECT * FROM main.xxx;" to copy
  ** the contents to the temporary database.
  */
  assert( (db->flags & SQLITE_Vacuum)==0 );
  db->flags |= SQLITE_Vacuum;
  rc = execExecSql(db, pzErrMsg,
      "SELECT 'INSERT INTO vacuum_db.' || quote(name) "
      "|| ' SELECT * FROM main.' || quote(name) || ';'"
      "FROM main.sqlite_master "
      "WHERE type = 'table' AND name!='sqlite_sequence' "
      "  AND coalesce(rootpage,1)>0"
  );
  assert( (db->flags & SQLITE_Vacuum)!=0 );
  db->flags &= ~SQLITE_Vacuum;
  if( rc!=SQLITE_OK ) goto end_of_vacuum;

  /* Copy over the sequence table
  */
  rc = execExecSql(db, pzErrMsg,
      "SELECT 'DELETE FROM vacuum_db.' || quote(name) || ';' "
      "FROM vacuum_db.sqlite_master WHERE name='sqlite_sequence' "

Changes to src/vdbe.c.

4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
....
4509
4510
4511
4512
4513
4514
4515

4516
4517
4518
4519
4520
4521
4522
  pC->cacheStatus = CACHE_STALE;
  if( pC->pCursor ){
    sqlite3BtreeClearCursor(pC->pCursor);
  }
  break;
}

/* Opcode: Last P1 P2 * * *
**
** The next use of the Rowid or Column or Prev instruction for P1 
** will refer to the last entry in the database table or index.
** If the table or index is empty and P2>0, then jump immediately to P2.
** If P2 is 0 or if the table or index is not empty, fall through
** to the following instruction.
**
................................................................................
  pCrsr = pC->pCursor;
  res = 0;
  assert( pCrsr!=0 );
  rc = sqlite3BtreeLast(pCrsr, &res);
  pC->nullRow = (u8)res;
  pC->deferredMoveto = 0;
  pC->cacheStatus = CACHE_STALE;

#ifdef SQLITE_DEBUG
  pC->seekOp = OP_Last;
#endif
  if( pOp->p2>0 ){
    VdbeBranchTaken(res!=0,2);
    if( res ) pc = pOp->p2 - 1;
  }







|







 







>







4482
4483
4484
4485
4486
4487
4488
4489
4490
4491
4492
4493
4494
4495
4496
....
4509
4510
4511
4512
4513
4514
4515
4516
4517
4518
4519
4520
4521
4522
4523
  pC->cacheStatus = CACHE_STALE;
  if( pC->pCursor ){
    sqlite3BtreeClearCursor(pC->pCursor);
  }
  break;
}

/* Opcode: Last P1 P2 P3 * *
**
** The next use of the Rowid or Column or Prev instruction for P1 
** will refer to the last entry in the database table or index.
** If the table or index is empty and P2>0, then jump immediately to P2.
** If P2 is 0 or if the table or index is not empty, fall through
** to the following instruction.
**
................................................................................
  pCrsr = pC->pCursor;
  res = 0;
  assert( pCrsr!=0 );
  rc = sqlite3BtreeLast(pCrsr, &res);
  pC->nullRow = (u8)res;
  pC->deferredMoveto = 0;
  pC->cacheStatus = CACHE_STALE;
  pC->seekResult = pOp->p3;
#ifdef SQLITE_DEBUG
  pC->seekOp = OP_Last;
#endif
  if( pOp->p2>0 ){
    VdbeBranchTaken(res!=0,2);
    if( res ) pc = pOp->p2 - 1;
  }

Changes to src/vdbe.h.

209
210
211
212
213
214
215

216
217
218
219
220
221
222
#ifndef SQLITE_OMIT_TRACE
  char *sqlite3VdbeExpandSql(Vdbe*, const char*);
#endif
int sqlite3MemCompare(const Mem*, const Mem*, const CollSeq*);

void sqlite3VdbeRecordUnpack(KeyInfo*,int,const void*,UnpackedRecord*);
int sqlite3VdbeRecordCompare(int,const void*,UnpackedRecord*);

UnpackedRecord *sqlite3VdbeAllocUnpackedRecord(KeyInfo *, char *, int, char **);

typedef int (*RecordCompare)(int,const void*,UnpackedRecord*);
RecordCompare sqlite3VdbeFindCompare(UnpackedRecord*);

#ifndef SQLITE_OMIT_TRIGGER
void sqlite3VdbeLinkSubProgram(Vdbe *, SubProgram *);







>







209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
#ifndef SQLITE_OMIT_TRACE
  char *sqlite3VdbeExpandSql(Vdbe*, const char*);
#endif
int sqlite3MemCompare(const Mem*, const Mem*, const CollSeq*);

void sqlite3VdbeRecordUnpack(KeyInfo*,int,const void*,UnpackedRecord*);
int sqlite3VdbeRecordCompare(int,const void*,UnpackedRecord*);
int sqlite3VdbeRecordCompareWithSkip(int, const void *, UnpackedRecord *, int);
UnpackedRecord *sqlite3VdbeAllocUnpackedRecord(KeyInfo *, char *, int, char **);

typedef int (*RecordCompare)(int,const void*,UnpackedRecord*);
RecordCompare sqlite3VdbeFindCompare(UnpackedRecord*);

#ifndef SQLITE_OMIT_TRIGGER
void sqlite3VdbeLinkSubProgram(Vdbe *, SubProgram *);

Changes to src/vdbeaux.c.

3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
....
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
....
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
....
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
** returned.
**
** If database corruption is discovered, set pPKey2->errCode to 
** SQLITE_CORRUPT and return 0. If an OOM error is encountered, 
** pPKey2->errCode is set to SQLITE_NOMEM and, if it is not NULL, the
** malloc-failed flag set on database handle (pPKey2->pKeyInfo->db).
*/
static int vdbeRecordCompareWithSkip(
  int nKey1, const void *pKey1,   /* Left key */
  UnpackedRecord *pPKey2,         /* Right key */
  int bSkip                       /* If true, skip the first field */
){
  u32 d1;                         /* Offset into aKey[] of next data element */
  int i;                          /* Index of next field to compare */
  u32 szHdr1;                     /* Size of record header in bytes */
................................................................................
  );
  return pPKey2->default_rc;
}
int sqlite3VdbeRecordCompare(
  int nKey1, const void *pKey1,   /* Left key */
  UnpackedRecord *pPKey2          /* Right key */
){
  return vdbeRecordCompareWithSkip(nKey1, pKey1, pPKey2, 0);
}


/*
** This function is an optimized version of sqlite3VdbeRecordCompare() 
** that (a) the first field of pPKey2 is an integer, and (b) the 
** size-of-header varint at the start of (pKey1/nKey1) fits in a single
................................................................................
  if( v>lhs ){
    res = pPKey2->r1;
  }else if( v<lhs ){
    res = pPKey2->r2;
  }else if( pPKey2->nField>1 ){
    /* The first fields of the two keys are equal. Compare the trailing 
    ** fields.  */
    res = vdbeRecordCompareWithSkip(nKey1, pKey1, pPKey2, 1);
  }else{
    /* The first fields of the two keys are equal and there are no trailing
    ** fields. Return pPKey2->default_rc in this case. */
    res = pPKey2->default_rc;
  }

  assert( vdbeRecordCompareDebug(nKey1, pKey1, pPKey2, res) );
................................................................................
    nCmp = MIN( pPKey2->aMem[0].n, nStr );
    res = memcmp(&aKey1[szHdr], pPKey2->aMem[0].z, nCmp);

    if( res==0 ){
      res = nStr - pPKey2->aMem[0].n;
      if( res==0 ){
        if( pPKey2->nField>1 ){
          res = vdbeRecordCompareWithSkip(nKey1, pKey1, pPKey2, 1);
        }else{
          res = pPKey2->default_rc;
        }
      }else if( res>0 ){
        res = pPKey2->r2;
      }else{
        res = pPKey2->r1;







|







 







|







 







|







 







|







3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
....
3767
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
....
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
....
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
3917
** returned.
**
** If database corruption is discovered, set pPKey2->errCode to 
** SQLITE_CORRUPT and return 0. If an OOM error is encountered, 
** pPKey2->errCode is set to SQLITE_NOMEM and, if it is not NULL, the
** malloc-failed flag set on database handle (pPKey2->pKeyInfo->db).
*/
int sqlite3VdbeRecordCompareWithSkip(
  int nKey1, const void *pKey1,   /* Left key */
  UnpackedRecord *pPKey2,         /* Right key */
  int bSkip                       /* If true, skip the first field */
){
  u32 d1;                         /* Offset into aKey[] of next data element */
  int i;                          /* Index of next field to compare */
  u32 szHdr1;                     /* Size of record header in bytes */
................................................................................
  );
  return pPKey2->default_rc;
}
int sqlite3VdbeRecordCompare(
  int nKey1, const void *pKey1,   /* Left key */
  UnpackedRecord *pPKey2          /* Right key */
){
  return sqlite3VdbeRecordCompareWithSkip(nKey1, pKey1, pPKey2, 0);
}


/*
** This function is an optimized version of sqlite3VdbeRecordCompare() 
** that (a) the first field of pPKey2 is an integer, and (b) the 
** size-of-header varint at the start of (pKey1/nKey1) fits in a single
................................................................................
  if( v>lhs ){
    res = pPKey2->r1;
  }else if( v<lhs ){
    res = pPKey2->r2;
  }else if( pPKey2->nField>1 ){
    /* The first fields of the two keys are equal. Compare the trailing 
    ** fields.  */
    res = sqlite3VdbeRecordCompareWithSkip(nKey1, pKey1, pPKey2, 1);
  }else{
    /* The first fields of the two keys are equal and there are no trailing
    ** fields. Return pPKey2->default_rc in this case. */
    res = pPKey2->default_rc;
  }

  assert( vdbeRecordCompareDebug(nKey1, pKey1, pPKey2, res) );
................................................................................
    nCmp = MIN( pPKey2->aMem[0].n, nStr );
    res = memcmp(&aKey1[szHdr], pPKey2->aMem[0].z, nCmp);

    if( res==0 ){
      res = nStr - pPKey2->aMem[0].n;
      if( res==0 ){
        if( pPKey2->nField>1 ){
          res = sqlite3VdbeRecordCompareWithSkip(nKey1, pKey1, pPKey2, 1);
        }else{
          res = pPKey2->default_rc;
        }
      }else if( res>0 ){
        res = pPKey2->r2;
      }else{
        res = pPKey2->r1;

Changes to src/vdbesort.c.

287
288
289
290
291
292
293

294
295
296
297
298
299
300

301
302
303

304
305
306
307
308
309
310
...
324
325
326
327
328
329
330

331
332



333
334
335
336
337
338
339
...
738
739
740
741
742
743
744


















745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760

761
762
763
764
765
766

767
768
769













































































































770
771
772
773
774
775
776
...
831
832
833
834
835
836
837


838

839
840

841
842
843
844
845
846
847
...
859
860
861
862
863
864
865






866
867
868
869
870
871
872
....
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201



1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212

1213
1214
1215
1216
1217













1218
1219
1220
1221
1222
1223
1224
....
1226
1227
1228
1229
1230
1231
1232



1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
....
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
....
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
....
1593
1594
1595
1596
1597
1598
1599










1600
1601
1602
1603
1604
1605
1606
....
1858
1859
1860
1861
1862
1863
1864


1865
1866
1867

1868
1869
1870
1871
1872
1873
1874
1875
....
2279
2280
2281
2282
2283
2284
2285





2286
2287
2288
2289
2290
2291
2292
**      to sqlite3ThreadJoin() is likely to block. Cases that are likely to
**      block provoke debugging output.
**
** In both cases, the effects of the main thread seeing (bDone==0) even
** after the thread has finished are not dire. So we don't worry about
** memory barriers and such here.
*/

struct SortSubtask {
  SQLiteThread *pThread;          /* Background thread, if any */
  int bDone;                      /* Set if thread is finished but not joined */
  VdbeSorter *pSorter;            /* Sorter that owns this sub-task */
  UnpackedRecord *pUnpacked;      /* Space to unpack a record */
  SorterList list;                /* List for thread to write to a PMA */
  int nPMA;                       /* Number of PMAs currently in file */

  SorterFile file;                /* Temp file for level-0 PMAs */
  SorterFile file2;               /* Space for other PMAs */
};


/*
** Main sorter structure. A single instance of this is allocated for each 
** sorter cursor created by the VDBE.
**
** mxKeysize:
**   As records are added to the sorter by calls to sqlite3VdbeSorterWrite(),
................................................................................
  SorterList list;                /* List of in-memory records */
  int iMemory;                    /* Offset of free space in list.aMemory */
  int nMemory;                    /* Size of list.aMemory allocation in bytes */
  u8 bUsePMA;                     /* True if one or more PMAs created */
  u8 bUseThreads;                 /* True to use background threads */
  u8 iPrev;                       /* Previous thread used to flush PMA */
  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.
................................................................................

  if( rc==SQLITE_OK ){
    rc = vdbePmaReaderNext(pReadr);
  }
  return rc;
}




















/*
** Compare key1 (buffer pKey1, size nKey1 bytes) with key2 (buffer pKey2, 
** size nKey2 bytes). Use (pTask->pKeyInfo) for the collation sequences
** used by the comparison. Return the result of the comparison.
**
** Before returning, object (pTask->pUnpacked) is populated with the
** unpacked version of key2. Or, if pKey2 is passed a NULL pointer, then it 
** is assumed that the (pTask->pUnpacked) structure already contains the 
** unpacked key to use as key2.
**
** If an OOM error is encountered, (pTask->pUnpacked->error_rc) is set
** to SQLITE_NOMEM.
*/
static int vdbeSorterCompare(
  SortSubtask *pTask,             /* Subtask context (for pKeyInfo) */

  const void *pKey1, int nKey1,   /* Left side of comparison */
  const void *pKey2, int nKey2    /* Right side of comparison */
){
  UnpackedRecord *r2 = pTask->pUnpacked;
  if( pKey2 ){
    sqlite3VdbeRecordUnpack(pTask->pSorter->pKeyInfo, nKey2, pKey2, r2);

  }
  return sqlite3VdbeRecordCompare(nKey1, pKey1, r2);
}














































































































/*
** Initialize the temporary index cursor just opened as a sorter cursor.
**
** Usually, the sorter module uses the value of (pCsr->pKeyInfo->nField)
** to determine the number of fields that should be compared from the
** records being sorted. However, if the value passed as argument nField
................................................................................
  pCsr->pSorter = pSorter;
  if( pSorter==0 ){
    rc = SQLITE_NOMEM;
  }else{
    pSorter->pKeyInfo = pKeyInfo = (KeyInfo*)((u8*)pSorter + sz);
    memcpy(pKeyInfo, pCsr->pKeyInfo, szKeyInfo);
    pKeyInfo->db = 0;


    if( nField && nWorker==0 ) pKeyInfo->nField = nField;

    pSorter->pgsz = pgsz = sqlite3BtreeGetPageSize(db->aDb[0].pBt);
    pSorter->nTask = nWorker + 1;

    pSorter->bUseThreads = (pSorter->nTask>1);
    pSorter->db = db;
    for(i=0; i<pSorter->nTask; i++){
      SortSubtask *pTask = &pSorter->aTask[i];
      pTask->pSorter = pSorter;
    }

................................................................................
      if( sqlite3GlobalConfig.pScratch==0 ){
        assert( pSorter->iMemory==0 );
        pSorter->nMemory = pgsz;
        pSorter->list.aMemory = (u8*)sqlite3Malloc(pgsz);
        if( !pSorter->list.aMemory ) rc = SQLITE_NOMEM;
      }
    }






  }

  return rc;
}
#undef nWorker   /* Defined at the top of this function */

/*
................................................................................
  SortSubtask *pTask,             /* Calling thread context */
  SorterRecord *p1,               /* First list to merge */
  SorterRecord *p2,               /* Second list to merge */
  SorterRecord **ppOut            /* OUT: Head of merged list */
){
  SorterRecord *pFinal = 0;
  SorterRecord **pp = &pFinal;
  void *pVal2 = p2 ? SRVAL(p2) : 0;

  while( p1 && p2 ){
    int res;
    res = vdbeSorterCompare(pTask, SRVAL(p1), p1->nVal, pVal2, p2->nVal);



    if( res<=0 ){
      *pp = p1;
      pp = &p1->u.pNext;
      p1 = p1->u.pNext;
      pVal2 = 0;
    }else{
      *pp = p2;
       pp = &p2->u.pNext;
      p2 = p2->u.pNext;
      if( p2==0 ) break;
      pVal2 = SRVAL(p2);

    }
  }
  *pp = p1 ? p1 : p2;
  *ppOut = pFinal;
}














/*
** Sort the linked list of records headed at pTask->pList. Return 
** SQLITE_OK if successful, or an SQLite error code (i.e. SQLITE_NOMEM) if 
** an error occurs.
*/
static int vdbeSorterSort(SortSubtask *pTask, SorterList *pList){
................................................................................
  SorterRecord **aSlot;
  SorterRecord *p;
  int rc;

  rc = vdbeSortAllocUnpacked(pTask);
  if( rc!=SQLITE_OK ) return rc;




  aSlot = (SorterRecord **)sqlite3MallocZero(64 * sizeof(SorterRecord *));
  if( !aSlot ){
    return SQLITE_NOMEM;
  }

  p = pList->pList;
  while( p ){
    SorterRecord *pNext;
    if( pList->aMemory ){
      if( (u8*)p==pList->aMemory ){
        pNext = 0;
      }else{
        assert( p->u.iNext<sqlite3MallocSize(pList->aMemory) );
................................................................................
  rc = vdbePmaReaderNext(&pMerger->aReadr[iPrev]);

  /* Update contents of aTree[] */
  if( rc==SQLITE_OK ){
    int i;                      /* Index of aTree[] to recalculate */
    PmaReader *pReadr1;         /* First PmaReader to compare */
    PmaReader *pReadr2;         /* Second PmaReader to compare */
    u8 *pKey2;                  /* To pReadr2->aKey, or 0 if record cached */

    /* Find the first two PmaReaders to compare. The one that was just
    ** advanced (iPrev) and the one next to it in the array.  */
    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
        );
      }

      /* If pReadr1 contained the smaller value, set aTree[i] to its index.
      ** Then set pReadr2 to the next PmaReader to compare to pReadr1. In this
      ** case there is no cache of pReadr2 in pTask->pUnpacked, so set
      ** pKey2 to point to the record belonging to pReadr2.
................................................................................
      ** If the two values were equal, then the value from the oldest
      ** PMA should be considered smaller. The VdbeSorter.aReadr[] array
      ** 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);
  }

................................................................................
  VdbeSorter *pSorter = pCsr->pSorter;
  int rc = SQLITE_OK;             /* Return Code */
  SorterRecord *pNew;             /* New list element */

  int bFlush;                     /* True to flush contents of memory to PMA */
  int nReq;                       /* Bytes of memory required */
  int nPMA;                       /* Bytes of PMA space required */











  assert( pSorter );

  /* Figure out whether or not the current contents of memory should be
  ** flushed to a PMA before continuing. If so, do so.
  **
  ** If using the single large allocation mode (pSorter->aMemory!=0), then
................................................................................
  p2 = &pMerger->aReadr[i2];

  if( p1->pFd==0 ){
    iRes = i2;
  }else if( p2->pFd==0 ){
    iRes = i1;
  }else{


    int res;
    assert( pMerger->pTask->pUnpacked!=0 );  /* from vdbeSortSubtaskMain() */
    res = vdbeSorterCompare(

        pMerger->pTask, p1->aKey, p1->nKey, p2->aKey, p2->nKey
    );
    if( res<=0 ){
      iRes = i1;
    }else{
      iRes = i2;
    }
  }
................................................................................
*/
static int vdbeSorterSetupMerge(VdbeSorter *pSorter){
  int rc;                         /* Return code */
  SortSubtask *pTask0 = &pSorter->aTask[0];
  MergeEngine *pMain = 0;
#if SQLITE_MAX_WORKER_THREADS
  sqlite3 *db = pTask0->pSorter->db;





#endif

  rc = vdbeSorterMergeTreeBuild(pSorter, &pMain);
  if( rc==SQLITE_OK ){
#if SQLITE_MAX_WORKER_THREADS
    assert( pSorter->bUseThreads==0 || pSorter->nTask>1 );
    if( pSorter->bUseThreads ){







>







>



>







 







>


>
>
>







 







>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>






|
|
|
|






>




|

>



>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>







 







>
>
|
>


>







 







>
>
>
>
>
>







 







|



|
>
>
>




<


|

<
<
>





>
>
>
>
>
>
>
>
>
>
>
>
>







 







>
>
>





<







 







|





<









|
|







 







|

|







 







>
>
>
>
>
>
>
>
>
>







 







>
>

|
<
>
|







 







>
>
>
>
>







287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
...
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
...
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
...
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
...
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
....
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354

1355
1356
1357
1358


1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
....
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400

1401
1402
1403
1404
1405
1406
1407
....
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619

1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
....
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
....
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
....
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039

2040
2041
2042
2043
2044
2045
2046
2047
2048
....
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
**      to sqlite3ThreadJoin() is likely to block. Cases that are likely to
**      block provoke debugging output.
**
** In both cases, the effects of the main thread seeing (bDone==0) even
** after the thread has finished are not dire. So we don't worry about
** memory barriers and such here.
*/
typedef int (*SorterCompare)(SortSubtask*,int*,const void*,int,const void*,int);
struct SortSubtask {
  SQLiteThread *pThread;          /* Background thread, if any */
  int bDone;                      /* Set if thread is finished but not joined */
  VdbeSorter *pSorter;            /* Sorter that owns this sub-task */
  UnpackedRecord *pUnpacked;      /* Space to unpack a record */
  SorterList list;                /* List for thread to write to a PMA */
  int nPMA;                       /* Number of PMAs currently in file */
  SorterCompare xCompare;         /* Compare function to use */
  SorterFile file;                /* Temp file for level-0 PMAs */
  SorterFile file2;               /* Space for other PMAs */
};


/*
** Main sorter structure. A single instance of this is allocated for each 
** sorter cursor created by the VDBE.
**
** mxKeysize:
**   As records are added to the sorter by calls to sqlite3VdbeSorterWrite(),
................................................................................
  SorterList list;                /* List of in-memory records */
  int iMemory;                    /* Offset of free space in list.aMemory */
  int nMemory;                    /* Size of list.aMemory allocation in bytes */
  u8 bUsePMA;                     /* True if one or more PMAs created */
  u8 bUseThreads;                 /* True to use background threads */
  u8 iPrev;                       /* Previous thread used to flush PMA */
  u8 nTask;                       /* Size of aTask[] array */
  u8 typeMask;
  SortSubtask aTask[1];           /* One or more subtasks */
};

#define SORTER_TYPE_INTEGER 0x01
#define SORTER_TYPE_TEXT    0x02

/*
** 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.
................................................................................

  if( rc==SQLITE_OK ){
    rc = vdbePmaReaderNext(pReadr);
  }
  return rc;
}

/*
** A version of vdbeSorterCompare() that assumes that it has already been
** determined that the first field of key1 is equal to the first field of 
** key2.
*/
static int vdbeSorterCompareTail(
  SortSubtask *pTask,             /* Subtask context (for pKeyInfo) */
  int *pbKey2Cached,              /* True if pTask->pUnpacked is pKey2 */
  const void *pKey1, int nKey1,   /* Left side of comparison */
  const void *pKey2, int nKey2    /* Right side of comparison */
){
  UnpackedRecord *r2 = pTask->pUnpacked;
  if( *pbKey2Cached==0 ){
    sqlite3VdbeRecordUnpack(pTask->pSorter->pKeyInfo, nKey2, pKey2, r2);
    *pbKey2Cached = 1;
  }
  return sqlite3VdbeRecordCompareWithSkip(nKey1, pKey1, r2, 1);
}

/*
** Compare key1 (buffer pKey1, size nKey1 bytes) with key2 (buffer pKey2, 
** size nKey2 bytes). Use (pTask->pKeyInfo) for the collation sequences
** used by the comparison. Return the result of the comparison.
**
** If IN/OUT parameter *pbKey2Cached is true when this function is called,
** it is assumed that (pTask->pUnpacked) contains the unpacked version
** of key2. If it is false, (pTask->pUnpacked) is populated with the unpacked
** version of key2 and *pbKey2Cached set to true before returning.
**
** If an OOM error is encountered, (pTask->pUnpacked->error_rc) is set
** to SQLITE_NOMEM.
*/
static int vdbeSorterCompare(
  SortSubtask *pTask,             /* Subtask context (for pKeyInfo) */
  int *pbKey2Cached,              /* True if pTask->pUnpacked is pKey2 */
  const void *pKey1, int nKey1,   /* Left side of comparison */
  const void *pKey2, int nKey2    /* Right side of comparison */
){
  UnpackedRecord *r2 = pTask->pUnpacked;
  if( !*pbKey2Cached ){
    sqlite3VdbeRecordUnpack(pTask->pSorter->pKeyInfo, nKey2, pKey2, r2);
    *pbKey2Cached = 1;
  }
  return sqlite3VdbeRecordCompare(nKey1, pKey1, r2);
}

/*
** A specially optimized version of vdbeSorterCompare() that assumes that
** the first field of each key is a TEXT value and that the collation
** sequence to compare them with is BINARY.
*/
static int vdbeSorterCompareText(
  SortSubtask *pTask,             /* Subtask context (for pKeyInfo) */
  int *pbKey2Cached,              /* True if pTask->pUnpacked is pKey2 */
  const void *pKey1, int nKey1,   /* Left side of comparison */
  const void *pKey2, int nKey2    /* Right side of comparison */
){
  const u8 * const p1 = (const u8 * const)pKey1;
  const u8 * const p2 = (const u8 * const)pKey2;
  const u8 * const v1 = &p1[ p1[0] ];   /* Pointer to value 1 */
  const u8 * const v2 = &p2[ p2[0] ];   /* Pointer to value 2 */

  int n1;
  int n2;
  int res;

  getVarint32(&p1[1], n1); n1 = (n1 - 13) / 2;
  getVarint32(&p2[1], n2); n2 = (n2 - 13) / 2;
  res = memcmp(v1, v2, MIN(n1, n2));
  if( res==0 ){
    res = n1 - n2;
  }

  if( res==0 ){
    if( pTask->pSorter->pKeyInfo->nField>1 ){
      res = vdbeSorterCompareTail(
          pTask, pbKey2Cached, pKey1, nKey1, pKey2, nKey2
      );
    }
  }else{
    if( pTask->pSorter->pKeyInfo->aSortOrder[0] ){
      res = res * -1;
    }
  }

  return res;
}

/*
** A specially optimized version of vdbeSorterCompare() that assumes that
** the first field of each key is an INTEGER value.
*/
static int vdbeSorterCompareInt(
  SortSubtask *pTask,             /* Subtask context (for pKeyInfo) */
  int *pbKey2Cached,              /* True if pTask->pUnpacked is pKey2 */
  const void *pKey1, int nKey1,   /* Left side of comparison */
  const void *pKey2, int nKey2    /* Right side of comparison */
){
  const u8 * const p1 = (const u8 * const)pKey1;
  const u8 * const p2 = (const u8 * const)pKey2;
  const int s1 = p1[1];                 /* Left hand serial type */
  const int s2 = p2[1];                 /* Right hand serial type */
  const u8 * const v1 = &p1[ p1[0] ];   /* Pointer to value 1 */
  const u8 * const v2 = &p2[ p2[0] ];   /* Pointer to value 2 */
  int res;                              /* Return value */

  assert( (s1>0 && s1<7) || s1==8 || s1==9 );
  assert( (s2>0 && s2<7) || s2==8 || s2==9 );

  if( s1>7 && s2>7 ){
    res = s1 - s2;
  }else{
    if( s1==s2 ){
      if( (*v1 ^ *v2) & 0x80 ){
        /* The two values have different signs */
        res = (*v1 & 0x80) ? -1 : +1;
      }else{
        /* The two values have the same sign. Compare using memcmp(). */
        static const u8 aLen[] = {0, 1, 2, 3, 4, 6, 8 };
        int i;
        res = 0;
        for(i=0; i<aLen[s1]; i++){
          if( (res = v1[i] - v2[i]) ) break;
        }
      }
    }else{
      if( s2>7 ){
        res = +1;
      }else if( s1>7 ){
        res = -1;
      }else{
        res = s1 - s2;
      }

      if( res>0 ){
        if( *v1 & 0x80 ) res = -1;
      }else if( res<0 ){
        if( *v2 & 0x80 ) res = +1;
      }
    }
  }

  if( res==0 ){
    if( pTask->pSorter->pKeyInfo->nField>1 ){
      res = vdbeSorterCompareTail(
          pTask, pbKey2Cached, pKey1, nKey1, pKey2, nKey2
      );
    }
  }else if( pTask->pSorter->pKeyInfo->aSortOrder[0] ){
    res = res * -1;
  }

  return res;
}

/*
** Initialize the temporary index cursor just opened as a sorter cursor.
**
** Usually, the sorter module uses the value of (pCsr->pKeyInfo->nField)
** to determine the number of fields that should be compared from the
** records being sorted. However, if the value passed as argument nField
................................................................................
  pCsr->pSorter = pSorter;
  if( pSorter==0 ){
    rc = SQLITE_NOMEM;
  }else{
    pSorter->pKeyInfo = pKeyInfo = (KeyInfo*)((u8*)pSorter + sz);
    memcpy(pKeyInfo, pCsr->pKeyInfo, szKeyInfo);
    pKeyInfo->db = 0;
    if( nField && nWorker==0 ){
      pKeyInfo->nXField += (pKeyInfo->nField - nField);
      pKeyInfo->nField = nField;
    }
    pSorter->pgsz = pgsz = sqlite3BtreeGetPageSize(db->aDb[0].pBt);
    pSorter->nTask = nWorker + 1;
    pSorter->iPrev = nWorker-1;
    pSorter->bUseThreads = (pSorter->nTask>1);
    pSorter->db = db;
    for(i=0; i<pSorter->nTask; i++){
      SortSubtask *pTask = &pSorter->aTask[i];
      pTask->pSorter = pSorter;
    }

................................................................................
      if( sqlite3GlobalConfig.pScratch==0 ){
        assert( pSorter->iMemory==0 );
        pSorter->nMemory = pgsz;
        pSorter->list.aMemory = (u8*)sqlite3Malloc(pgsz);
        if( !pSorter->list.aMemory ) rc = SQLITE_NOMEM;
      }
    }

    if( (pKeyInfo->nField+pKeyInfo->nXField)<13 
     && (pKeyInfo->aColl[0]==0 || pKeyInfo->aColl[0]==db->pDfltColl)
    ){
      pSorter->typeMask = SORTER_TYPE_INTEGER | SORTER_TYPE_TEXT;
    }
  }

  return rc;
}
#undef nWorker   /* Defined at the top of this function */

/*
................................................................................
  SortSubtask *pTask,             /* Calling thread context */
  SorterRecord *p1,               /* First list to merge */
  SorterRecord *p2,               /* Second list to merge */
  SorterRecord **ppOut            /* OUT: Head of merged list */
){
  SorterRecord *pFinal = 0;
  SorterRecord **pp = &pFinal;
  int bCached = 0;

  while( p1 && p2 ){
    int res;
    res = pTask->xCompare(
        pTask, &bCached, SRVAL(p1), p1->nVal, SRVAL(p2), p2->nVal
    );

    if( res<=0 ){
      *pp = p1;
      pp = &p1->u.pNext;
      p1 = p1->u.pNext;

    }else{
      *pp = p2;
      pp = &p2->u.pNext;
      p2 = p2->u.pNext;


      bCached = 0;
    }
  }
  *pp = p1 ? p1 : p2;
  *ppOut = pFinal;
}

/*
** Return the SorterCompare function to compare values collected by the
** sorter object passed as the only argument.
*/
static SorterCompare vdbeSorterGetCompare(VdbeSorter *p){
  if( p->typeMask==SORTER_TYPE_INTEGER ){
    return vdbeSorterCompareInt;
  }else if( p->typeMask==SORTER_TYPE_TEXT ){
    return vdbeSorterCompareText; 
  }
  return vdbeSorterCompare;
}

/*
** Sort the linked list of records headed at pTask->pList. Return 
** SQLITE_OK if successful, or an SQLite error code (i.e. SQLITE_NOMEM) if 
** an error occurs.
*/
static int vdbeSorterSort(SortSubtask *pTask, SorterList *pList){
................................................................................
  SorterRecord **aSlot;
  SorterRecord *p;
  int rc;

  rc = vdbeSortAllocUnpacked(pTask);
  if( rc!=SQLITE_OK ) return rc;

  p = pList->pList;
  pTask->xCompare = vdbeSorterGetCompare(pTask->pSorter);

  aSlot = (SorterRecord **)sqlite3MallocZero(64 * sizeof(SorterRecord *));
  if( !aSlot ){
    return SQLITE_NOMEM;
  }


  while( p ){
    SorterRecord *pNext;
    if( pList->aMemory ){
      if( (u8*)p==pList->aMemory ){
        pNext = 0;
      }else{
        assert( p->u.iNext<sqlite3MallocSize(pList->aMemory) );
................................................................................
  rc = vdbePmaReaderNext(&pMerger->aReadr[iPrev]);

  /* Update contents of aTree[] */
  if( rc==SQLITE_OK ){
    int i;                      /* Index of aTree[] to recalculate */
    PmaReader *pReadr1;         /* First PmaReader to compare */
    PmaReader *pReadr2;         /* Second PmaReader to compare */
    int bCached = 0;

    /* Find the first two PmaReaders to compare. The one that was just
    ** advanced (iPrev) and the one next to it in the array.  */
    pReadr1 = &pMerger->aReadr[(iPrev & 0xFFFE)];
    pReadr2 = &pMerger->aReadr[(iPrev | 0x0001)];


    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 = pTask->xCompare(pTask, &bCached,
            pReadr1->aKey, pReadr1->nKey, pReadr2->aKey, pReadr2->nKey
        );
      }

      /* If pReadr1 contained the smaller value, set aTree[i] to its index.
      ** Then set pReadr2 to the next PmaReader to compare to pReadr1. In this
      ** case there is no cache of pReadr2 in pTask->pUnpacked, so set
      ** pKey2 to point to the record belonging to pReadr2.
................................................................................
      ** If the two values were equal, then the value from the oldest
      ** PMA should be considered smaller. The VdbeSorter.aReadr[] array
      ** 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] ];
        bCached = 0;
      }else{
        if( pReadr1->pFd ) bCached = 0;
        pMerger->aTree[i] = (int)(pReadr2 - pMerger->aReadr);
        pReadr1 = &pMerger->aReadr[ pMerger->aTree[i ^ 0x0001] ];
      }
    }
    *pbEof = (pMerger->aReadr[pMerger->aTree[1]].pFd==0);
  }

................................................................................
  VdbeSorter *pSorter = pCsr->pSorter;
  int rc = SQLITE_OK;             /* Return Code */
  SorterRecord *pNew;             /* New list element */

  int bFlush;                     /* True to flush contents of memory to PMA */
  int nReq;                       /* Bytes of memory required */
  int nPMA;                       /* Bytes of PMA space required */
  int t;                          /* serial type of first record field */

  getVarint32((const u8*)&pVal->z[1], t);
  if( t>0 && t<10 && t!=7 ){
    pSorter->typeMask &= SORTER_TYPE_INTEGER;
  }else if( t>10 && (t & 0x01) ){
    pSorter->typeMask &= SORTER_TYPE_TEXT;
  }else{
    pSorter->typeMask = 0;
  }

  assert( pSorter );

  /* Figure out whether or not the current contents of memory should be
  ** flushed to a PMA before continuing. If so, do so.
  **
  ** If using the single large allocation mode (pSorter->aMemory!=0), then
................................................................................
  p2 = &pMerger->aReadr[i2];

  if( p1->pFd==0 ){
    iRes = i2;
  }else if( p2->pFd==0 ){
    iRes = i1;
  }else{
    SortSubtask *pTask = pMerger->pTask;
    int bCached = 0;
    int res;
    assert( pTask->pUnpacked!=0 );  /* from vdbeSortSubtaskMain() */

    res = pTask->xCompare(
        pTask, &bCached, p1->aKey, p1->nKey, p2->aKey, p2->nKey
    );
    if( res<=0 ){
      iRes = i1;
    }else{
      iRes = i2;
    }
  }
................................................................................
*/
static int vdbeSorterSetupMerge(VdbeSorter *pSorter){
  int rc;                         /* Return code */
  SortSubtask *pTask0 = &pSorter->aTask[0];
  MergeEngine *pMain = 0;
#if SQLITE_MAX_WORKER_THREADS
  sqlite3 *db = pTask0->pSorter->db;
  int i;
  SorterCompare xCompare = vdbeSorterGetCompare(pSorter);
  for(i=0; i<pSorter->nTask; i++){
    pSorter->aTask[i].xCompare = xCompare;
  }
#endif

  rc = vdbeSorterMergeTreeBuild(pSorter, &pMain);
  if( rc==SQLITE_OK ){
#if SQLITE_MAX_WORKER_THREADS
    assert( pSorter->bUseThreads==0 || pSorter->nTask>1 );
    if( pSorter->bUseThreads ){