executor: fix the behavior when index join meet prefix index (pingcap…

…#11081)

executor/builder.go

@@ -1731,6 +1731,13 @@ func (b *executorBuilder) buildIndexLookUpJoin(v *plannercore.PhysicalIndexJoin)
 	if defaultValues == nil {
 		defaultValues = make([]types.Datum, len(innerTypes))
 	}
+	hasPrefixCol := false
+	for _, l := range v.IdxColLens {
+		if l != types.UnspecifiedLength {
+			hasPrefixCol = true
+			break
+		}
+	}
 	e := &IndexLookUpJoin{
 		baseExecutor: newBaseExecutor(b.ctx, v.Schema(), v.ExplainID(), outerExec),
 		outerCtx: outerCtx{
@@ -1740,6 +1747,8 @@ func (b *executorBuilder) buildIndexLookUpJoin(v *plannercore.PhysicalIndexJoin)
 		innerCtx: innerCtx{
 			readerBuilder: &dataReaderBuilder{Plan: innerPlan, executorBuilder: b},
 			rowTypes:      innerTypes,
+			colLens:       v.IdxColLens,
+			hasPrefixCol:  hasPrefixCol,
 		},
 		workerWg:      new(sync.WaitGroup),
 		joiner:        newJoiner(b.ctx, v.JoinType, v.OuterIndex == 1, defaultValues, v.OtherConditions, leftTypes, rightTypes),

executor/index_lookup_join.go

@@ -89,6 +89,8 @@ type innerCtx struct {
 	readerBuilder *dataReaderBuilder
 	rowTypes      []*types.FieldType
 	keyCols       []int
+	colLens       []int
+	hasPrefixCol  bool
 }
 
 type lookUpJoinTask struct {
@@ -490,6 +492,16 @@ func (iw *innerWorker) constructLookupContent(task *lookUpJoinTask) ([]*indexJoi
 		}
 		// Store the encoded lookup key in chunk, so we can use it to lookup the matched inners directly.
 		task.encodedLookUpKeys.AppendBytes(0, keyBuf)
+		if iw.hasPrefixCol {
+			for i := range iw.outerCtx.keyCols {
+				// If it's a prefix column. Try to fix it.
+				if iw.colLens[i] != types.UnspecifiedLength {
+					ranger.CutDatumByPrefixLen(&dLookUpKey[i], iw.colLens[i], iw.rowTypes[iw.keyCols[i]])
+				}
+			}
+			// dLookUpKey is sorted and deduplicated at sortAndDedupLookUpContents.
+			// So we don't need to do it here.
+		}
 		lookUpContents = append(lookUpContents, &indexJoinLookUpContent{keys: dLookUpKey, row: task.outerResult.GetRow(i)})
 	}
 

executor/index_lookup_join_test.go

@@ -152,3 +152,11 @@ func (s *testSuite) TestIndexJoinOverflow(c *C) {
 	tk.MustExec(`create table t2(a int unsigned, index idx(a));`)
 	tk.MustQuery(`select /*+ TIDB_INLJ(t2) */ * from t1 join t2 on t1.a = t2.a;`).Check(testkit.Rows())
 }
+
+func (s *testSuite2) TestIssue11061(c *C) {
+	tk := testkit.NewTestKitWithInit(c, s.store)
+	tk.MustExec("drop table if exists t1, t2")
+	tk.MustExec("create table t1(c varchar(30), index ix_c(c(10)))")
+	tk.MustExec("insert into t1 (c) values('7_chars'), ('13_characters')")
+	tk.MustQuery("SELECT /*+ TIDB_INLJ(t1) */ SUM(LENGTH(c)) FROM t1 WHERE c IN (SELECT t1.c FROM t1)").Check(testkit.Rows("20"))
+}

planner/core/exhaust_physical_plans.go

@@ -321,8 +321,15 @@ func joinKeysMatchIndex(keys, indexCols []*expression.Column, colLengths []int)
 
 // When inner plan is TableReader, the parameter `ranges` will be nil. Because pk only have one column. So all of its range
 // is generated during execution time.
-func (p *LogicalJoin) constructIndexJoin(prop *property.PhysicalProperty, outerIdx int, innerPlan PhysicalPlan,
-	ranges []*ranger.Range, keyOff2IdxOff []int, compareFilters *ColWithCmpFuncManager) []PhysicalPlan {
+func (p *LogicalJoin) constructIndexJoin(
+	prop *property.PhysicalProperty,
+	outerIdx int,
+	innerPlan PhysicalPlan,
+	ranges []*ranger.Range,
+	keyOff2IdxOff []int,
+	lens []int,
+	compareFilters *ColWithCmpFuncManager,
+) []PhysicalPlan {
 	joinType := p.JoinType
 	outerSchema := p.children[outerIdx].Schema()
 	var (
@@ -372,6 +379,7 @@ func (p *LogicalJoin) constructIndexJoin(prop *property.PhysicalProperty, outerI
 		DefaultValues:   p.DefaultValues,
 		innerPlan:       innerPlan,
 		KeyOff2IdxOff:   newKeyOff,
+		IdxColLens:      lens,
 		Ranges:          ranges,
 		CompareFilters:  compareFilters,
 	}.Init(p.ctx, p.stats.ScaleByExpectCnt(prop.ExpectedCnt), chReqProps...)
@@ -430,7 +438,7 @@ func (p *LogicalJoin) getIndexJoinByOuterIdx(prop *property.PhysicalProperty, ou
 			innerPlan := p.constructInnerTableScan(ds, pkCol, outerJoinKeys, us)
 			// Since the primary key means one value corresponding to exact one row, this will always be a no worse one
 			// comparing to other index.
-			return p.constructIndexJoin(prop, outerIdx, innerPlan, nil, keyOff2IdxOff, nil)
+			return p.constructIndexJoin(prop, outerIdx, innerPlan, nil, keyOff2IdxOff, nil, nil)
 		}
 	}
 	helper := &indexJoinBuildHelper{join: p}
@@ -454,10 +462,10 @@ func (p *LogicalJoin) getIndexJoinByOuterIdx(prop *property.PhysicalProperty, ou
 				keyOff2IdxOff[keyOff] = idxOff
 			}
 		}
-		idxCols, _ := expression.IndexInfo2Cols(ds.schema.Columns, helper.chosenIndexInfo)
+		idxCols, lens := expression.IndexInfo2Cols(ds.schema.Columns, helper.chosenIndexInfo)
 		rangeInfo := helper.buildRangeDecidedByInformation(idxCols, outerJoinKeys)
 		innerPlan := p.constructInnerIndexScan(ds, helper.chosenIndexInfo, helper.chosenRemained, outerJoinKeys, us, rangeInfo)
-		return p.constructIndexJoin(prop, outerIdx, innerPlan, helper.chosenRanges, keyOff2IdxOff, helper.lastColManager)
+		return p.constructIndexJoin(prop, outerIdx, innerPlan, helper.chosenRanges, keyOff2IdxOff, lens, helper.lastColManager)
 	}
 	return nil
 }

planner/core/physical_plans.go

@@ -232,6 +232,8 @@ type PhysicalIndexJoin struct {
 	Ranges []*ranger.Range
 	// KeyOff2IdxOff maps the offsets in join key to the offsets in the index.
 	KeyOff2IdxOff []int
+	// IdxColLens stores the length of each index column.
+	IdxColLens []int
 	// CompareFilters stores the filters for last column if those filters need to be evaluated during execution.
 	// e.g. select * from t where t.a = t1.a and t.b > t1.b and t.b < t1.b+10
 	//      If there's index(t.a, t.b). All the filters can be used to construct index range but t.b > t1.b and t.b < t1.b=10

util/ranger/ranger.go

@@ -263,10 +263,10 @@ func buildColumnRange(accessConditions []expression.Expression, sc *stmtctx.Stat
 	}
 	if colLen != types.UnspecifiedLength {
 		for _, ran := range ranges {
-			if fixRangeDatum(&ran.LowVal[0], colLen, tp) {
+			if CutDatumByPrefixLen(&ran.LowVal[0], colLen, tp) {
 				ran.LowExclude = false
 			}
-			if fixRangeDatum(&ran.HighVal[0], colLen, tp) {
+			if CutDatumByPrefixLen(&ran.HighVal[0], colLen, tp) {
 				ran.HighExclude = false
 			}
 		}
@@ -425,17 +425,17 @@ func fixPrefixColRange(ranges []*Range, lengths []int, tp []*types.FieldType) bo
 	for _, ran := range ranges {
 		lowTail := len(ran.LowVal) - 1
 		for i := 0; i < lowTail; i++ {
-			fixRangeDatum(&ran.LowVal[i], lengths[i], tp[i])
+			CutDatumByPrefixLen(&ran.LowVal[i], lengths[i], tp[i])
 		}
-		lowCut := fixRangeDatum(&ran.LowVal[lowTail], lengths[lowTail], tp[lowTail])
+		lowCut := CutDatumByPrefixLen(&ran.LowVal[lowTail], lengths[lowTail], tp[lowTail])
 		if lowCut {
 			ran.LowExclude = false
 		}
 		highTail := len(ran.HighVal) - 1
 		for i := 0; i < highTail; i++ {
-			fixRangeDatum(&ran.HighVal[i], lengths[i], tp[i])
+			CutDatumByPrefixLen(&ran.HighVal[i], lengths[i], tp[i])
 		}
-		highCut := fixRangeDatum(&ran.HighVal[highTail], lengths[highTail], tp[highTail])
+		highCut := CutDatumByPrefixLen(&ran.HighVal[highTail], lengths[highTail], tp[highTail])
 		if highCut {
 			ran.HighExclude = false
 		}
@@ -444,9 +444,9 @@ func fixPrefixColRange(ranges []*Range, lengths []int, tp []*types.FieldType) bo
 	return hasCut
 }
 
-func fixRangeDatum(v *types.Datum, length int, tp *types.FieldType) bool {
-	// If this column is prefix and the prefix length is smaller than the range, cut it.
-	// In case of UTF8, prefix should be cut by characters rather than bytes
+// CutDatumByPrefixLen cuts the datum according to the prefix length.
+// If it's UTF8 encoded, we will cut it by characters rather than bytes.
+func CutDatumByPrefixLen(v *types.Datum, length int, tp *types.FieldType) bool {
 	if v.Kind() == types.KindString || v.Kind() == types.KindBytes {
 		colCharset := tp.Charset
 		colValue := v.GetBytes()