/ Check-in [b9c695e8]
Login

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

Overview
Comment:Add the OP_SorterColumns opcode - an experiment in using a special case opcode to decode the Sorter output rather than the generic OP_Column. This might be faster. And with further work, it could eventually eliminate the need for OP_OpenPseudo.
Downloads: Tarball | ZIP archive | SQL archive
Timelines: family | ancestors | descendants | both | OP_SorterColumns
Files: files | file ages | folders
SHA1: b9c695e8859bb9972a9890bf2ccf341c1282ab77
User & Date: drh 2014-10-13 01:23:51
Context
2014-10-13
12:30
Use OP_SorterColumns in aggregate queries. Remove OPFLAG_CLEARCACHE. Closed-Leaf check-in: 134e65c0 user: drh tags: OP_SorterColumns
01:23
Add the OP_SorterColumns opcode - an experiment in using a special case opcode to decode the Sorter output rather than the generic OP_Column. This might be faster. And with further work, it could eventually eliminate the need for OP_OpenPseudo. check-in: b9c695e8 user: drh tags: OP_SorterColumns
2014-10-12
22:37
Remove the VdbeCursor.lastRowid cache of the current rowid, since maintaining the correct cache value uses more CPU cycles than just recomputing the rowid on the occasions when it is actually needed. Replace it with the VdbeCursor.aOffset field which used to be computed from VdbeCursor.aType when needed. Saves 100 bytes of code space and runs 0.2% faster. check-in: 91384a7d user: drh tags: trunk
Changes
Hide Diffs Unified Diffs Ignore Whitespace Patch

Changes to src/select.c.

1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
....
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231

1232
1233
1234
1235
1236
1237
1238
  int nColumn,      /* Number of columns of data */
  SelectDest *pDest /* Write the sorted results here */
){
  Vdbe *v = pParse->pVdbe;                     /* The prepared statement */
  int addrBreak = sqlite3VdbeMakeLabel(v);     /* Jump here to exit loop */
  int addrContinue = sqlite3VdbeMakeLabel(v);  /* Jump here for next cycle */
  int addr;
  int addrOnce = 0;
  int iTab;
  ExprList *pOrderBy = pSort->pOrderBy;
  int eDest = pDest->eDest;
  int iParm = pDest->iSDParm;
  int regRow;
  int regRowid;
  int nKey;
  int iSortTab;                   /* Sorter cursor to read from */
  int nSortData;                  /* Trailing values to read from sorter */
  u8 p5;                          /* p5 parameter for 1st OP_Column */
  int i;
  int bSeq;                       /* True if sorter record includes seq. no. */
#ifdef SQLITE_ENABLE_EXPLAIN_COMMENTS
  struct ExprList_item *aOutEx = p->pEList->a;
#endif

  if( pSort->labelBkOut ){
    sqlite3VdbeAddOp2(v, OP_Gosub, pSort->regReturn, pSort->labelBkOut);
    sqlite3VdbeAddOp2(v, OP_Goto, 0, addrBreak);
................................................................................
    regRowid = sqlite3GetTempReg(pParse);
    regRow = sqlite3GetTempReg(pParse);
    nSortData = 1;
  }
  nKey = pOrderBy->nExpr - pSort->nOBSat;
  if( pSort->sortFlags & SORTFLAG_UseSorter ){
    int regSortOut = ++pParse->nMem;
    iSortTab = pParse->nTab++;
    if( pSort->labelBkOut ){
      addrOnce = sqlite3CodeOnce(pParse); VdbeCoverage(v);
    }
    sqlite3VdbeAddOp3(v, OP_OpenPseudo, iSortTab, regSortOut, nKey+1+nSortData);
    if( addrOnce ) sqlite3VdbeJumpHere(v, addrOnce);
    addr = 1 + sqlite3VdbeAddOp2(v, OP_SorterSort, iTab, addrBreak);
    VdbeCoverage(v);
    codeOffset(v, p->iOffset, addrContinue);
    sqlite3VdbeAddOp2(v, OP_SorterData, iTab, regSortOut);
    p5 = OPFLAG_CLEARCACHE;
    bSeq = 0;
  }else{
    addr = 1 + sqlite3VdbeAddOp2(v, OP_Sort, iTab, addrBreak); VdbeCoverage(v);
    codeOffset(v, p->iOffset, addrContinue);
    iSortTab = iTab;
    p5 = 0;
    bSeq = 1;
  }
  for(i=0; i<nSortData; i++){
    sqlite3VdbeAddOp3(v, OP_Column, iSortTab, nKey+bSeq+i, regRow+i);
    if( i==0 ) sqlite3VdbeChangeP5(v, p5);
    VdbeComment((v, "%s", aOutEx[i].zName ? aOutEx[i].zName : aOutEx[i].zSpan));

  }
  switch( eDest ){
    case SRT_Table:
    case SRT_EphemTab: {
      testcase( eDest==SRT_Table );
      testcase( eDest==SRT_EphemTab );
      sqlite3VdbeAddOp2(v, OP_NewRowid, iParm, regRowid);







<







<

<

<







 







<
<
<
<
<
<




|
<



<
<
<
<
|
|
<
|
>







1167
1168
1169
1170
1171
1172
1173

1174
1175
1176
1177
1178
1179
1180

1181

1182

1183
1184
1185
1186
1187
1188
1189
....
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
  int nColumn,      /* Number of columns of data */
  SelectDest *pDest /* Write the sorted results here */
){
  Vdbe *v = pParse->pVdbe;                     /* The prepared statement */
  int addrBreak = sqlite3VdbeMakeLabel(v);     /* Jump here to exit loop */
  int addrContinue = sqlite3VdbeMakeLabel(v);  /* Jump here for next cycle */
  int addr;

  int iTab;
  ExprList *pOrderBy = pSort->pOrderBy;
  int eDest = pDest->eDest;
  int iParm = pDest->iSDParm;
  int regRow;
  int regRowid;
  int nKey;

  int nSortData;                  /* Trailing values to read from sorter */

  int i;

#ifdef SQLITE_ENABLE_EXPLAIN_COMMENTS
  struct ExprList_item *aOutEx = p->pEList->a;
#endif

  if( pSort->labelBkOut ){
    sqlite3VdbeAddOp2(v, OP_Gosub, pSort->regReturn, pSort->labelBkOut);
    sqlite3VdbeAddOp2(v, OP_Goto, 0, addrBreak);
................................................................................
    regRowid = sqlite3GetTempReg(pParse);
    regRow = sqlite3GetTempReg(pParse);
    nSortData = 1;
  }
  nKey = pOrderBy->nExpr - pSort->nOBSat;
  if( pSort->sortFlags & SORTFLAG_UseSorter ){
    int regSortOut = ++pParse->nMem;






    addr = 1 + sqlite3VdbeAddOp2(v, OP_SorterSort, iTab, addrBreak);
    VdbeCoverage(v);
    codeOffset(v, p->iOffset, addrContinue);
    sqlite3VdbeAddOp2(v, OP_SorterData, iTab, regSortOut);
    sqlite3VdbeAddOp4Int(v, OP_SorterColumns, nKey, nSortData, regRow, regSortOut);

  }else{
    addr = 1 + sqlite3VdbeAddOp2(v, OP_Sort, iTab, addrBreak); VdbeCoverage(v);
    codeOffset(v, p->iOffset, addrContinue);




    for(i=0; i<nSortData; i++){
      sqlite3VdbeAddOp3(v, OP_Column, iTab, nKey+i+1, regRow+i);

      VdbeComment((v, "%s", aOutEx[i].zName ? aOutEx[i].zName : aOutEx[i].zSpan));
    }
  }
  switch( eDest ){
    case SRT_Table:
    case SRT_EphemTab: {
      testcase( eDest==SRT_Table );
      testcase( eDest==SRT_EphemTab );
      sqlite3VdbeAddOp2(v, OP_NewRowid, iParm, regRowid);

Changes to src/vdbe.c.

2284
2285
2286
2287
2288
2289
2290

2291
2292
2293
2294
2295
2296
2297
....
4247
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
















































4263
4264
4265
4266
4267
4268
4269
  pCrsr = pC->pCursor;
  assert( pCrsr!=0 || pC->pseudoTableReg>0 ); /* pCrsr NULL on PseudoTables */
  assert( pCrsr!=0 || pC->nullRow );          /* pC->nullRow on PseudoTables */

  /* If the cursor cache is stale, bring it up-to-date */
  rc = sqlite3VdbeCursorMoveto(pC);
  if( rc ) goto abort_due_to_error;

  if( pC->cacheStatus!=p->cacheCtr || (pOp->p5&OPFLAG_CLEARCACHE)!=0 ){
    if( pC->nullRow ){
      if( pCrsr==0 ){
        assert( pC->pseudoTableReg>0 );
        pReg = &aMem[pC->pseudoTableReg];
        assert( pReg->flags & MEM_Blob );
        assert( memIsValid(pReg) );
................................................................................

/* Opcode: SorterData P1 P2 * * *
** Synopsis: r[P2]=data
**
** Write into register P2 the current sorter data for sorter cursor P1.
*/
case OP_SorterData: {
  VdbeCursor *pC;

  pOut = &aMem[pOp->p2];
  pC = p->apCsr[pOp->p1];
  assert( isSorter(pC) );
  rc = sqlite3VdbeSorterRowkey(pC, pOut);
  assert( rc!=SQLITE_OK || (pOut->flags & MEM_Blob) );
  break;
}

















































/* Opcode: RowData P1 P2 * * *
** Synopsis: r[P2]=data
**
** Write into register P2 the complete row data for cursor P1.
** There is no interpretation of the data.  
** It is just copied onto the P2 register exactly as 







>







 







|








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







2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
....
4248
4249
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
4309
4310
4311
4312
4313
4314
4315
4316
4317
4318
  pCrsr = pC->pCursor;
  assert( pCrsr!=0 || pC->pseudoTableReg>0 ); /* pCrsr NULL on PseudoTables */
  assert( pCrsr!=0 || pC->nullRow );          /* pC->nullRow on PseudoTables */

  /* If the cursor cache is stale, bring it up-to-date */
  rc = sqlite3VdbeCursorMoveto(pC);
  if( rc ) goto abort_due_to_error;
  assert( (pOp->p5&OPFLAG_CLEARCACHE)==0 || aOp[pc-1].opcode==OP_SorterData );
  if( pC->cacheStatus!=p->cacheCtr || (pOp->p5&OPFLAG_CLEARCACHE)!=0 ){
    if( pC->nullRow ){
      if( pCrsr==0 ){
        assert( pC->pseudoTableReg>0 );
        pReg = &aMem[pC->pseudoTableReg];
        assert( pReg->flags & MEM_Blob );
        assert( memIsValid(pReg) );
................................................................................

/* Opcode: SorterData P1 P2 * * *
** Synopsis: r[P2]=data
**
** Write into register P2 the current sorter data for sorter cursor P1.
*/
case OP_SorterData: {
  VdbeCursor *pC;         /* Sorting cursor defined by P1 */

  pOut = &aMem[pOp->p2];
  pC = p->apCsr[pOp->p1];
  assert( isSorter(pC) );
  rc = sqlite3VdbeSorterRowkey(pC, pOut);
  assert( rc!=SQLITE_OK || (pOut->flags & MEM_Blob) );
  break;
}

/* Opcode: SorterColumns P1 P2 P3 P4 *
** Synopsis: r[P3@P2]=decode(r[P4])
**
** The P4 register contains a record that has just come out of a sorter.
** Decode columns P1 through P1+P2-1 into registers P3..P3+P2-1.
**
** This opcode is much faster than multiple calls to Column since it 
** does not need to deal with corrupt record detection or default values
** or any of the other complications associated with a record read
** from disk.
*/
case OP_SorterColumns: {
  Mem *pDest;             /* Register P3 output register */
  Mem *pLast;             /* Register P3+P2-1 */
  u32 serial_type;        /* Serial type of a column value */
  u32 idx;                /* Index into the record header */
  u32 d;                  /* Index into the data of the record */
  int nSkip;              /* Number of initial columns to skip */
  const u8 *aKey;         /* Complete text of the record */

  assert( pOp->p4type==P4_INT32 );
  assert( pOp->p4.i>0 && pOp->p4.i<=(p->nMem - p->nCursor) );
  assert( pOp->p3>pOp->p4.i || pOp->p3+pOp->p2<=pOp->p4.i );
  assert( pOp->p1>=0 );
  assert( pOp->p2>0 );
  assert( aMem[pOp->p4.i].flags & MEM_Blob );
  aKey = (const u8*)aMem[pOp->p4.i].z;
  pDest = &aMem[pOp->p3];
  pLast = &pDest[pOp->p2-1];
  idx = getVarint32(aKey, d);
  nSkip = pOp->p1;
  while( nSkip-- ){
    assert( d<=aMem[pOp->p4.i].n );
    idx += getVarint32(&aKey[idx], serial_type);
    d += sqlite3VdbeSerialTypeLen(serial_type);
  }
  do{
    assert( d<=aMem[pOp->p4.i].n );
    idx += getVarint32(&aKey[idx], serial_type);
    if( VdbeMemDynamic(pDest) ) sqlite3VdbeMemSetNull(pDest);
    d += sqlite3VdbeSerialGet(&aKey[d], serial_type, pDest);
    pDest->enc = encoding;
    REGISTER_TRACE((int)(pDest-aMem), pDest);
    pDest++;
  }while( pDest<=pLast );
  break;
}

/* Opcode: RowData P1 P2 * * *
** Synopsis: r[P2]=data
**
** Write into register P2 the complete row data for cursor P1.
** There is no interpretation of the data.  
** It is just copied onto the P2 register exactly as