props: add methods to EquivGroups for use in FuncDepSet

This commit adds several new methods along with unit tests to `EquivGroups`
to prepare its use in tracking equivalencies in `FuncDepSet`.

Informs cockroachdb#83963

Release note: None

pkg/sql/opt/norm/join_funcs.go

@@ -460,7 +460,7 @@ func (c *CustomFuncs) GetEquivColsWithEquivTypeWithEquivGroups(
 	}
 
 	// Compute all equivalent columns.
-	eqCols := equivSet.Group(col)
+	eqCols := equivSet.GroupForCol(col)
 
 	eqCols.ForEach(func(i opt.ColumnID) {
 		// Only include columns that have the same type as col.

pkg/sql/opt/props/BUILD.bazel

@@ -63,6 +63,7 @@ go_test(
         "//pkg/sql/sem/tree",
         "//pkg/sql/sem/volatility",
         "//pkg/sql/types",
+        "//pkg/util/buildutil",
         "//pkg/util/randutil",
         "@com_github_cockroachdb_errors//:errors",
         "@com_github_stretchr_testify//require",

pkg/sql/opt/props/equiv_set.go

@@ -30,26 +30,149 @@ func (eq *EquivGroups) Reset() {
 	eq.groups = eq.groups[:0]
 }
 
-// Add adds the given equivalent columns to the EquivGroups. If possible, the
-// columns are added to an existing group. Otherwise, a new one is created.
-func (eq *EquivGroups) Add(equivCols opt.ColSet) {
+// Empty returns true if the set stores no equalities.
+func (eq *EquivGroups) Empty() bool {
 	if buildutil.CrdbTestBuild {
 		defer eq.verify()
 	}
+	return len(eq.groups) == 0
+}
+
+// GroupCount returns the number of equiv groups stored in the set.
+func (eq *EquivGroups) GroupCount() int {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	return len(eq.groups)
+}
+
+// Group returns the equiv group at the given index. The returned ColSet should
+// be considered immutable. The index must be less than GroupCount().
+func (eq *EquivGroups) Group(idx int) opt.ColSet {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	if idx >= len(eq.groups) {
+		panic(errors.AssertionFailedf("invalid equiv group index %d", idx))
+	}
+	return eq.groups[idx]
+}
+
+// GroupForCol returns the group of columns equivalent to the given column. It
+// returns the empty set if no such group exists. The returned should not be
+// mutated without being copied first.
+func (eq *EquivGroups) GroupForCol(col opt.ColumnID) opt.ColSet {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	for i := range eq.groups {
+		if eq.groups[i].Contains(col) {
+			return eq.groups[i]
+		}
+	}
+	return opt.ColSet{}
+}
+
+// ContainsCol returns true if the given column is contained in any of the equiv
+// groups (it will be in at most one group).
+func (eq *EquivGroups) ContainsCol(col opt.ColumnID) bool {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	for i := range eq.groups {
+		if eq.groups[i].Contains(col) {
+			return true
+		}
+	}
+	return false
+}
+
+// AreColsEquiv indicates whether the given columns are equivalent.
+func (eq *EquivGroups) AreColsEquiv(left, right opt.ColumnID) bool {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	if left == right {
+		return true
+	}
+	for i := range eq.groups {
+		if eq.groups[i].Contains(left) {
+			return eq.groups[i].Contains(right)
+		}
+		if eq.groups[i].Contains(right) {
+			return eq.groups[i].Contains(left)
+		}
+	}
+	return false
+}
+
+// AreAllColsEquiv returns true if all columns in the given set are equivalent
+// to all others in the set.
+func (eq *EquivGroups) AreAllColsEquiv(cols opt.ColSet) bool {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	if cols.Len() <= 1 {
+		return true
+	}
+	for i := range eq.groups {
+		if eq.groups[i].Intersects(cols) {
+			return cols.SubsetOf(eq.groups[i])
+		}
+	}
+	return false
+}
+
+// ComputeEquivClosureNoCopy returns the equivalence closure of the given
+// columns. Note that the given ColSet is mutated and returned directly.
+func (eq *EquivGroups) ComputeEquivClosureNoCopy(cols opt.ColSet) opt.ColSet {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	for i := range eq.groups {
+		if eq.groups[i].Intersects(cols) {
+			cols.UnionWith(eq.groups[i])
+			if cols.Len() == eq.groups[i].Len() {
+				// Since we just took the union, equal lengths means all columns in cols
+				// were within the same equivalence group, so we can short-circuit.
+				break
+			}
+		}
+	}
+	return cols
+}
+
+// AddNoCopy adds the given equivalent columns to the EquivGroups. If possible,
+// the columns are added to an existing group. Otherwise, a new one is created.
+// NOTE: the given ColSet may be added to the EquivGroups without being copied,
+// so it must be considered immutable after it is passed to addNoCopy.
+func (eq *EquivGroups) AddNoCopy(equivCols opt.ColSet) {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	if equivCols.Len() <= 1 {
+		// This is a trivial equivalence.
+		return
+	}
 	// Attempt to add the equivalence to an existing group.
 	for i := range eq.groups {
 		if eq.groups[i].Intersects(equivCols) {
 			if equivCols.SubsetOf(eq.groups[i]) {
-				// No-op
+				// The equivalence is already contained in the set.
 				return
 			}
-			eq.groups[i].UnionWith(equivCols)
+			if eq.groups[i].SubsetOf(equivCols) {
+				// Avoid the copy.
+				eq.groups[i] = equivCols
+			} else {
+				eq.groups[i] = eq.groups[i].Union(equivCols)
+			}
 			eq.tryMergeGroups(i)
 			return
 		}
 	}
 	// Make a new equivalence group.
-	eq.groups = append(eq.groups, equivCols.Copy())
+	eq.groups = append(eq.groups, equivCols)
 }
 
 // AddFromFDs adds all equivalence relations from the given FuncDepSet to the
@@ -61,49 +184,155 @@ func (eq *EquivGroups) AddFromFDs(fdset *FuncDepSet) {
 	for i := range fdset.deps {
 		fd := &fdset.deps[i]
 		if fd.equiv {
-			eq.Add(fd.from.Union(fd.to))
+			eq.AddNoCopy(fd.from.Union(fd.to))
 		}
 	}
 }
 
-// AreColsEquiv indicates whether the given columns are equivalent.
-func (eq *EquivGroups) AreColsEquiv(left, right opt.ColumnID) bool {
+// TranslateColsStrict remaps the column IDs of each equiv group according to
+// the given "from" and "to" lists. It requires that all columns in each group
+// are present in the "from" list, and that the "from" and "to" lists are the
+// same length.
+func (eq *EquivGroups) TranslateColsStrict(fromCols, toCols opt.ColList) {
 	if buildutil.CrdbTestBuild {
 		defer eq.verify()
 	}
+	// It is possible that the same column shows up more than once in either of
+	// the lists. In other words, a column can map to more than one column, and
+	// two different columns can map to the same column. The former is handled
+	// by TranslateColSetStrict, and may add a column to an equiv group. The
+	// latter can merge two equiv groups, so we need to handle it here.
+	var seenCols, dupCols opt.ColSet
+	for _, toCol := range toCols {
+		if seenCols.Contains(toCol) && !dupCols.Contains(toCol) {
+			var equiv opt.ColSet
+			for i, fromCol := range fromCols {
+				if toCols[i] == toCol {
+					equiv.Add(fromCol)
+				}
+			}
+			eq.AddNoCopy(equiv)
+			dupCols.Add(toCol)
+		}
+		seenCols.Add(toCol)
+	}
 	for i := range eq.groups {
-		if eq.groups[i].Contains(left) {
-			return eq.groups[i].Contains(right)
+		eq.groups[i] = opt.TranslateColSetStrict(eq.groups[i], fromCols, toCols)
+	}
+	// Handle the case when multiple "in" columns map to the same "out" column,
+	// which could result in removal of an equiv group.
+	eq.removeTrivialGroups()
+}
+
+// ProjectCols removes all columns from the EquivGroups that are not in the
+// given ColSet, removing equiv groups that become empty.
+func (eq *EquivGroups) ProjectCols(cols opt.ColSet) {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	for i := range eq.groups {
+		if !eq.groups[i].SubsetOf(cols) {
+			eq.groups[i] = eq.groups[i].Intersection(cols)
 		}
-		if eq.groups[i].Contains(right) {
-			return eq.groups[i].Contains(left)
+	}
+	eq.removeTrivialGroups()
+}
+
+// PartitionBy divides the equiv groups according to the given columns. If an
+// equiv group intersects the given ColSet but is not a subset, it is split into
+// the intersection and difference with the given ColSet. Ex:
+//
+//	eq := [(1-3), (4-8), (9-12)]
+//	eq.PartitionBy(1,5,6)
+//	eq == [(2,3), (4,7,8), (5,6), (9-12)]
+//
+// * In the example, the (1-3) group is split into (1) and (2,3). Since the (1)
+// group only has a single column, it is discarded as a trivial equivalence.
+// * The (4-8) group is split into (4,8) and (5,6). Since both subsets have at
+// least two columns, both are kept in the EquivGroups.
+// * Finally, the (9-12) group does not intersect the given cols, and so is not
+// split.
+func (eq *EquivGroups) PartitionBy(cols opt.ColSet) {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	for i := len(eq.groups) - 1; i >= 0; i-- {
+		if eq.groups[i].Intersects(cols) && !eq.groups[i].SubsetOf(cols) {
+			left, right := eq.groups[i].Intersection(cols), eq.groups[i].Difference(cols)
+			eq.groups[i] = opt.ColSet{}
+			if left.Len() > 1 {
+				eq.groups = append(eq.groups, left)
+			}
+			if right.Len() > 1 {
+				eq.groups = append(eq.groups, right)
+			}
 		}
 	}
-	return false
+	eq.removeTrivialGroups()
 }
 
-// Group returns the group of columns equivalent to the given column. It
-// returns the empty set if no such group exists. The returned should not be
-// mutated without being copied first.
-func (eq *EquivGroups) Group(col opt.ColumnID) opt.ColSet {
+// CopyFrom copies the given EquivGroups into this EquivGroups, replacing any
+// existing data.
+func (eq *EquivGroups) CopyFrom(other *EquivGroups) {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	eq.Reset()
+	eq.AppendFromDisjoint(other)
+}
+
+// AppendFromDisjoint unions the equiv groups from the given EquivGroups with
+// this one, assuming the groups are disjoint.
+func (eq *EquivGroups) AppendFromDisjoint(other *EquivGroups) {
+	if buildutil.CrdbTestBuild {
+		other.verify()
+		defer eq.verify()
+	}
+	neededCap := len(eq.groups) + len(other.groups)
+	if cap(eq.groups) < neededCap {
+		// Make sure to copy the old equiv groups into the new slice.
+		newGroups := make([]opt.ColSet, len(eq.groups), neededCap)
+		copy(newGroups, eq.groups)
+		eq.groups = newGroups
+	}
+	// There is no need to deep-copy the equiv groups, since they are never
+	// modified in-place.
+	eq.groups = append(eq.groups, other.groups...)
+}
+
+func (eq *EquivGroups) String() string {
+	if buildutil.CrdbTestBuild {
+		defer eq.verify()
+	}
+	ret := "["
 	for i := range eq.groups {
-		if eq.groups[i].Contains(col) {
-			return eq.groups[i]
+		if i > 0 {
+			ret += ", "
 		}
+		ret += eq.groups[i].String()
 	}
-	return opt.ColSet{}
+	return ret + "]"
 }
 
 // tryMergeGroups attempts to merge the equality group at the given index with
 // any of the *following* groups. If a group can be merged, it is removed after
 // its columns are added to the given group.
 func (eq *EquivGroups) tryMergeGroups(idx int) {
-	if buildutil.CrdbTestBuild {
-		defer eq.verify()
-	}
 	for i := len(eq.groups) - 1; i > idx; i-- {
 		if eq.groups[idx].Intersects(eq.groups[i]) {
-			eq.groups[idx].UnionWith(eq.groups[i])
+			eq.groups[idx] = eq.groups[idx].Union(eq.groups[i])
+			eq.groups[i] = eq.groups[len(eq.groups)-1]
+			eq.groups[len(eq.groups)-1] = opt.ColSet{}
+			eq.groups = eq.groups[:len(eq.groups)-1]
+		}
+	}
+}
+
+// removeTrivialGroups removes groups with zero or one columns, which may be
+// added by methods like makePartition.
+func (eq *EquivGroups) removeTrivialGroups() {
+	for i := len(eq.groups) - 1; i >= 0; i-- {
+		if eq.groups[i].Len() <= 1 {
 			eq.groups[i] = eq.groups[len(eq.groups)-1]
 			eq.groups[len(eq.groups)-1] = opt.ColSet{}
 			eq.groups = eq.groups[:len(eq.groups)-1]
@@ -123,14 +352,3 @@ func (eq *EquivGroups) verify() {
 		seen.UnionWith(group)
 	}
 }
-
-func (eq *EquivGroups) String() string {
-	ret := "["
-	for i := range eq.groups {
-		if i > 0 {
-			ret += ", "
-		}
-		ret += eq.groups[i].String()
-	}
-	return ret + "]"
-}