SQLite

Check-in [79ff36b717]
Login

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

Overview
Comment:Remove the use of htonl() in the previous check-in due to linkage issues. Add the get2byteAligned() macro and use it for access to the cell offsets on btree pages for about a 1% performance gain.
Downloads: Tarball | ZIP archive
Timelines: family | ancestors | descendants | both | bswap-functions
Files: files | file ages | folders
SHA1: 79ff36b7170c9e7e7a9935c8b9d1665867771087
User & Date: drh 2015-06-30 13:28:18.237
Context
2015-06-30
14:01
Only use __builtin_bswap16() with GCC 4.8 and later. (Closed-Leaf check-in: ce8177e3e6 user: drh tags: bswap-functions)
13:28
Remove the use of htonl() in the previous check-in due to linkage issues. Add the get2byteAligned() macro and use it for access to the cell offsets on btree pages for about a 1% performance gain. (check-in: 79ff36b717 user: drh tags: bswap-functions)
12:47
Make use of htonl() and __builtin_bswap32() for faster implementations of sqlite3Get4byte() and sqlite3Put4byte(). (check-in: bc27ebd7f7 user: drh tags: bswap-functions)
Changes
Side-by-Side Diff Ignore Whitespace Patch
Changes to src/btree.c.
︙
967
968
969
970
971
972
973
974

975
976

977
978
979
980
981
982
983

967
968
969
970
971
972
973

974
975

976
977
978
979
980
981
982
983








-
+

-
+








**
** findCellPastPtr() does the same except it skips past the initial
** 4-byte child pointer found on interior pages, if there is one.
**
** 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)])))
  ((P)->aData + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))
#define findCellPastPtr(P,I) \
  ((P)->aDataOfst + ((P)->maskPage & get2byte(&(P)->aCellIdx[2*(I)])))
  ((P)->aDataOfst + ((P)->maskPage & get2byteAligned(&(P)->aCellIdx[2*(I)])))


/*
** 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.
︙
1750
1751
1752
1753
1754
1755
1756
1757

1758
1759
1760
1761
1762
1763
1764

1750
1751
1752
1753
1754
1755
1756

1757
1758
1759
1760
1761
1762
1763
1764








-
+








    iCellLast = usableSize - 4;
    if( pBt->db->flags & SQLITE_CellSizeCk ){
      int i;            /* Index into the cell pointer array */
      int sz;           /* Size of a cell */

      if( !pPage->leaf ) iCellLast--;
      for(i=0; i<pPage->nCell; i++){
        pc = get2byte(&data[cellOffset+i*2]);
        pc = get2byteAligned(&data[cellOffset+i*2]);
        testcase( pc==iCellFirst );
        testcase( pc==iCellLast );
        if( pc<iCellFirst || pc>iCellLast ){
          return SQLITE_CORRUPT_BKPT;
        }
        sz = pPage->xCellSize(pPage, &data[pc]);
        testcase( pc+sz==usableSize );
︙
6621
6622
6623
6624
6625
6626
6627
6628

6629
6630
6631
6632
6633
6634
6635

6621
6622
6623
6624
6625
6626
6627

6628
6629
6630
6631
6632
6633
6634
6635








-
+









  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 = pCArray->apCell[i+iNew];
    int iOff = get2byte(&pPg->aCellIdx[i*2]);
    int iOff = get2byteAligned(&pPg->aCellIdx[i*2]);
    if( pCell>=aData && pCell<&aData[pPg->pBt->usableSize] ){
      pCell = &pTmp[pCell - aData];
    }
    assert( 0==memcmp(pCell, &aData[iOff],
            pCArray->pRef->xCellSize(pCArray->pRef, pCArray->apCell[i+iNew])) );
  }
#endif
︙
7123
7124
7125
7126
7127
7128
7129
7130

7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142
7143

7144
7145
7146
7147
7148
7149
7150

7123
7124
7125
7126
7127
7128
7129

7130
7131
7132
7133
7134
7135
7136
7137
7138
7139
7140
7141
7142

7143
7144
7145
7146
7147
7148
7149
7150








-
+












-
+








    ** 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));
        b.apCell[b.nCell] = aData + (maskPage & get2byteAligned(piCell));
        piCell += 2;
        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++;
      }
    }
    piEnd = aData + pOld->cellOffset + 2*pOld->nCell;
    while( piCell<piEnd ){
      assert( b.nCell<nMaxCells );
      b.apCell[b.nCell] = aData + (maskPage & get2byte(piCell));
      b.apCell[b.nCell] = aData + (maskPage & get2byteAligned(piCell));
      piCell += 2;
      b.nCell++;
    }

    cntOld[i] = b.nCell;
    if( i<nOld-1 && !leafData){
      u16 sz = (u16)szNew[i];
︙
9101
9102
9103
9104
9105
9106
9107
9108

9109
9110
9111
9112
9113
9114
9115

9101
9102
9103
9104
9105
9106
9107

9108
9109
9110
9111
9112
9113
9114
9115








-
+








    nCell = get2byte(&data[hdr+3]);
    /* EVIDENCE-OF: R-23882-45353 The cell pointer array of a b-tree page
    ** immediately follows the b-tree page header. */
    cellStart = hdr + 12 - 4*pPage->leaf;
    /* EVIDENCE-OF: R-02776-14802 The cell pointer array consists of K 2-byte
    ** integer offsets to the cell contents. */
    for(i=0; i<nCell; i++){
      int pc = get2byte(&data[cellStart+i*2]);
      int pc = get2byteAligned(&data[cellStart+i*2]);
      u32 size = 65536;
      if( pc<=usableSize-4 ){
        size = pPage->xCellSize(pPage, &data[pc]);
      }
      if( (int)(pc+size-1)>=usableSize ){
        pCheck->zPfx = 0;
        checkAppendMsg(pCheck,
︙
Changes to src/btreeInt.h.
︙
687
688
689
690
691
692
693














687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706








+
+
+
+
+
+
+
+
+
+
+
+
+

/*
** Routines to read or write a two- and four-byte big-endian integer values.
*/
#define get2byte(x)   ((x)[0]<<8 | (x)[1])
#define put2byte(p,v) ((p)[0] = (u8)((v)>>8), (p)[1] = (u8)(v))
#define get4byte sqlite3Get4byte
#define put4byte sqlite3Put4byte

/*
** get2byteAligned(), unlike get2byte(), requires that its argument point to a
** two-byte aligned address.  get2bytea() is only used for accessing the
** cell addresses in a btree header.
*/
#if SQLITE_BYTEORDER==4321
# define get2byteAligned(x)  (*(u16*)(x))
#elif SQLITE_BYTEORDER==1234 && defined(__GNUC__)
# define get2byteAligned(x)  __builtin_bswap16(*(u16*)(x))
#else
# define get2byteAligned(x)  ((x)[0]<<8 | (x)[1])
#endif
Changes to src/util.c.
︙
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110

1078
1079
1080
1081
1082
1083
1084




1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096



1097
1098
1099
1100
1101
1102
1103








-
-
-
-












-
-
-








** Read or write a four-byte big-endian integer value.
*/
u32 sqlite3Get4byte(const u8 *p){
#if SQLITE_BYTEORDER==4321
  u32 x;
  memcpy(&x,p,4);
  return x;
#elif defined(_MSC_VER)
  u32 x;
  memcpy(&x,p,4);
  return htonl(x);
#elif SQLITE_BYTEORDER==1234 && defined(__GNUC__)
  u32 x;
  memcpy(&x,p,4);
  return __builtin_bswap32(x);
#else
  testcase( p[0]&0x80 );
  return ((unsigned)p[0]<<24) | (p[1]<<16) | (p[2]<<8) | p[3];
#endif
}
void sqlite3Put4byte(unsigned char *p, u32 v){
#if SQLITE_BYTEORDER==4321
  memcpy(p,&v,4);
#elif defined(_MSC_VER)
  u32 x = htonl(v);
  memcpy(&x,p,4);
#elif SQLITE_BYTEORDER==1234 && defined(__GNUC__)
  u32 x = __builtin_bswap32(v);
  memcpy(p,&x,4);
#else
  p[0] = (u8)(v>>24);
  p[1] = (u8)(v>>16);
  p[2] = (u8)(v>>8);
︙