SQLite

Check-in [0f9bb90100]
Login

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

Overview
Comment:Fix a query planner problem that only occurs when covering-index-scan is disabled. Fix to tests whose output changed due to the new and more aggressive ORDER BY optimization.
Downloads: Tarball | ZIP archive
Timelines: family | ancestors | descendants | both | qp-enhancements
Files: files | file ages | folders
SHA1: 0f9bb90100aa304a7f28023ca4173e68b445e8bd
User & Date: drh 2012-10-03 12:38:19.551
Context
2012-10-03
12:56
Update the query planner to recognize more cases where ORDER BY clauses can be optimized out. Add test cases to verify correct behavior of the ORDER BY optimization when the covering-index-scan optimization is disabled. Fix a harmless compiler warning in the TCL interface. (check-in: 956e4d7f89 user: drh tags: trunk)
12:38
Fix a query planner problem that only occurs when covering-index-scan is disabled. Fix to tests whose output changed due to the new and more aggressive ORDER BY optimization. (Closed-Leaf check-in: 0f9bb90100 user: drh tags: qp-enhancements)
00:25
Further attempts to optimize out unnecessary ORDER BY clauses. (check-in: 6744d9a37f user: drh tags: qp-enhancements)
Changes
Unified Diff Ignore Whitespace Patch
Changes to src/where.c.
︙︙
3362
3363
3364
3365
3366
3367
3368

3369
3370
3371
3372
3373
3374
3375

      ** the cost function to err on the side of choosing an index over
      ** choosing a full scan.  This 4x full-scan penalty is an arguable
      ** decision and one which we expect to revisit in the future.  But
      ** it seems to be working well enough at the moment.
      */
      pc.rCost = aiRowEst[0]*4;
      pc.plan.wsFlags &= ~WHERE_IDX_ONLY;

    }else{
      log10N = estLog(aiRowEst[0]);
      pc.rCost = pc.plan.nRow;
      if( pIdx ){
        if( bLookup ){
          /* For an index lookup followed by a table lookup:
          **    nInMul index searches to find the start of each index range








>








3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
3375
3376

      ** the cost function to err on the side of choosing an index over
      ** choosing a full scan.  This 4x full-scan penalty is an arguable
      ** decision and one which we expect to revisit in the future.  But
      ** it seems to be working well enough at the moment.
      */
      pc.rCost = aiRowEst[0]*4;
      pc.plan.wsFlags &= ~WHERE_IDX_ONLY;
      if( pIdx ) pc.plan.wsFlags &= ~WHERE_ORDERED;
    }else{
      log10N = estLog(aiRowEst[0]);
      pc.rCost = pc.plan.nRow;
      if( pIdx ){
        if( bLookup ){
          /* For an index lookup followed by a table lookup:
          **    nInMul index searches to find the start of each index range
︙︙
Changes to test/where.test.
︙︙
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103

    CREATE TABLE t8(a INTEGER PRIMARY KEY, b TEXT UNIQUE);
    INSERT INTO t8 VALUES(1,'one');
    INSERT INTO t8 VALUES(4,'four');
  }
  cksort {
    SELECT x.a || '/' || y.a FROM t8 x, t8 y ORDER BY x.a, y.b
  } 
} {1/4 1/1 4/4 4/1 sort}
do_test where-14.2 {
  cksort {
    SELECT x.a || '/' || y.a FROM t8 x, t8 y ORDER BY x.a, y.b DESC
  } 
} {1/1 1/4 4/1 4/4 sort}
do_test where-14.3 {
  cksort {
    SELECT x.a || '/' || y.a FROM t8 x, t8 y ORDER BY x.a, x.b
  } 
} {1/4 1/1 4/4 4/1 nosort}
do_test where-14.4 {
  cksort {








|




|








1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103

    CREATE TABLE t8(a INTEGER PRIMARY KEY, b TEXT UNIQUE);
    INSERT INTO t8 VALUES(1,'one');
    INSERT INTO t8 VALUES(4,'four');
  }
  cksort {
    SELECT x.a || '/' || y.a FROM t8 x, t8 y ORDER BY x.a, y.b
  } 
} {1/4 1/1 4/4 4/1 nosort}
do_test where-14.2 {
  cksort {
    SELECT x.a || '/' || y.a FROM t8 x, t8 y ORDER BY x.a, y.b DESC
  } 
} {1/1 1/4 4/1 4/4 nosort}
do_test where-14.3 {
  cksort {
    SELECT x.a || '/' || y.a FROM t8 x, t8 y ORDER BY x.a, x.b
  } 
} {1/4 1/1 4/4 4/1 nosort}
do_test where-14.4 {
  cksort {
︙︙