// makeEqualityPredicate constructs a joinPredicate object for joins. The join
// condition includes equality between numMergedEqualityColumns columns,
// specified by leftColNames and rightColNames.
func makeEqualityPredicate(
left, right *dataSourceInfo,
leftColNames, rightColNames parser.NameList,
numMergedEqualityColumns int,
concatInfos *dataSourceInfo,
) (resPred *joinPredicate, info *dataSourceInfo, err error) {
if len(leftColNames) != len(rightColNames) {
panic(fmt.Errorf("left columns' length %q doesn't match right columns' length %q in EqualityPredicate",
len(leftColNames), len(rightColNames)))
}
if len(leftColNames) < numMergedEqualityColumns {
panic(fmt.Errorf("cannot merge %d columns, only %d columns to compare", numMergedEqualityColumns, len(leftColNames)))
}
// Prepare the arrays populated below.
cmpOps := make([]func(*parser.EvalContext, parser.Datum, parser.Datum) (parser.DBool, error), len(leftColNames))
leftEqualityIndices := make([]int, len(leftColNames))
rightEqualityIndices := make([]int, len(rightColNames))
// usedLeft represents the list of indices that participate in the
// equality predicate. They are collected in order to determine
// below which columns remain after the equality; this is used
// only when merging result columns.
var usedLeft, usedRight []int
var columns ResultColumns
if numMergedEqualityColumns > 0 {
usedLeft = make([]int, len(left.sourceColumns))
for i := range usedLeft {
usedLeft[i] = invalidColIdx
}
usedRight = make([]int, len(right.sourceColumns))
for i := range usedRight {
usedRight[i] = invalidColIdx
}
nResultColumns := len(left.sourceColumns) + len(right.sourceColumns) - numMergedEqualityColumns
columns = make(ResultColumns, 0, nResultColumns)
}
// Find out which columns are involved in EqualityPredicate.
for i := range leftColNames {
leftColName := leftColNames[i].Normalize()
rightColName := rightColNames[i].Normalize()
// Find the column name on the left.
leftIdx, leftType, err := pickUsingColumn(left.sourceColumns, leftColName, "left")
if err != nil {
return nil, nil, err
}
// Find the column name on the right.
rightIdx, rightType, err := pickUsingColumn(right.sourceColumns, rightColName, "right")
if err != nil {
return nil, nil, err
}
// Remember the indices.
leftEqualityIndices[i] = leftIdx
rightEqualityIndices[i] = rightIdx
// Memoize the comparison function.
fn, found := parser.FindEqualComparisonFunction(leftType, rightType)
if !found {
return nil, nil, fmt.Errorf("JOIN/USING types %s for left column %s and %s for right column %s cannot be matched",
leftType, leftColName, rightType, rightColName)
}
cmpOps[i] = fn
if i < numMergedEqualityColumns {
usedLeft[leftIdx] = i
usedRight[rightIdx] = i
// Merged columns come first in the results.
columns = append(columns, left.sourceColumns[leftIdx])
}
}
// Now, prepare/complete the metadata for the result columns.
// The structure of the join data source results is like this:
// - first, all the equality/USING columns;
// - then all the left columns,
// - then all the right columns,
// The duplicate columns appended after the equality/USING columns
// are hidden so that they are invisible to star expansion, but
// not omitted so that they can still be selected separately.
// Finish collecting the column definitions from the left and
// right data sources.
for i, c := range left.sourceColumns {
if usedLeft != nil && usedLeft[i] != invalidColIdx {
c.hidden = true
}
columns = append(columns, c)
}
for i, c := range right.sourceColumns {
if usedRight != nil && usedRight[i] != invalidColIdx {
c.hidden = true
}
columns = append(columns, c)
}
// Compute the mappings from table aliases to column sets from
// both sides into a new alias-columnset mapping for the result
// rows. We need to be extra careful about the aliases
// for the anonymous table, which needs to be merged.
aliases := make(sourceAliases, 0, len(left.sourceAliases)+len(right.sourceAliases))
collectAliases := func(sourceAliases sourceAliases, offset int) {
for _, alias := range sourceAliases {
if alias.name == anonymousTable {
continue
}
newRange := make([]int, len(alias.columnRange))
for i, colIdx := range alias.columnRange {
newRange[i] = colIdx + offset
}
aliases = append(aliases, sourceAlias{name: alias.name, columnRange: newRange})
}
}
collectAliases(left.sourceAliases, numMergedEqualityColumns)
collectAliases(right.sourceAliases, numMergedEqualityColumns+len(left.sourceColumns))
anonymousAlias := sourceAlias{name: anonymousTable, columnRange: nil}
var hiddenLeftNames, hiddenRightNames []string
// All the merged columns at the beginning belong to the
// anonymous data source.
for i := 0; i < numMergedEqualityColumns; i++ {
anonymousAlias.columnRange = append(anonymousAlias.columnRange, i)
hiddenLeftNames = append(hiddenLeftNames, parser.ReNormalizeName(left.sourceColumns[i].Name))
hiddenRightNames = append(hiddenRightNames, parser.ReNormalizeName(right.sourceColumns[i].Name))
}
// Now collect the other table-less columns into the anonymous data
// source, but hide (skip) those that are already merged.
collectAnonymousAliases := func(
sourceAliases sourceAliases, hiddenNames []string, cols ResultColumns, offset int,
) {
for _, alias := range sourceAliases {
if alias.name != anonymousTable {
continue
}
for _, colIdx := range alias.columnRange {
isHidden := false
for _, hiddenName := range hiddenNames {
if parser.ReNormalizeName(cols[colIdx].Name) == hiddenName {
isHidden = true
break
}
}
if !isHidden {
anonymousAlias.columnRange = append(anonymousAlias.columnRange, colIdx+offset)
}
}
}
}
collectAnonymousAliases(left.sourceAliases, hiddenLeftNames, left.sourceColumns,
numMergedEqualityColumns)
collectAnonymousAliases(right.sourceAliases, hiddenRightNames, right.sourceColumns,
numMergedEqualityColumns+len(left.sourceColumns))
if anonymousAlias.columnRange != nil {
aliases = append(aliases, anonymousAlias)
}
info = &dataSourceInfo{
sourceColumns: columns,
sourceAliases: aliases,
}
pred := &joinPredicate{
numLeftCols: len(left.sourceColumns),
numRightCols: len(right.sourceColumns),
leftColNames: leftColNames,
rightColNames: rightColNames,
numMergedEqualityColumns: numMergedEqualityColumns,
cmpFunctions: cmpOps,
leftEqualityIndices: leftEqualityIndices,
rightEqualityIndices: rightEqualityIndices,
info: info,
}
// We must initialize the indexed var helper in all cases, even when
// there is no on condition, so that getNeededColumns() does not get
// confused.
pred.iVarHelper = parser.MakeIndexedVarHelper(pred, len(columns))
return pred, info, nil
}
// tryAddEqualityFilter attempts to turn the given filter expression into
// an equality predicate. It returns true iff the transformation succeeds.
func (p *joinPredicate) tryAddEqualityFilter(filter parser.Expr, left, right *dataSourceInfo) bool {
c, ok := filter.(*parser.ComparisonExpr)
if !ok || c.Operator != parser.EQ {
return false
}
lhs, ok := c.Left.(*parser.IndexedVar)
if !ok {
return false
}
rhs, ok := c.Right.(*parser.IndexedVar)
if !ok {
return false
}
sourceBoundary := p.numMergedEqualityColumns + len(left.sourceColumns)
if (lhs.Idx >= sourceBoundary && rhs.Idx >= sourceBoundary) ||
(lhs.Idx < sourceBoundary && rhs.Idx < sourceBoundary) {
// Both variables are on the same side of the join (e.g. `a JOIN b ON a.x = a.y`).
return false
}
if lhs.Idx > rhs.Idx {
lhs, rhs = rhs, lhs
}
// At this point we have an equality, so we can add it to the list
// of equality columns.
// To do this we must be a bit careful: the expression contains
// IndexedVars, and the column indices at this point will refer to
// the full column set of the joinPredicate, including the
// merged columns.
leftColIdx := lhs.Idx - p.numMergedEqualityColumns
rightColIdx := rhs.Idx - len(left.sourceColumns) - p.numMergedEqualityColumns
// Also, we will want to avoid redundant equality checks.
for i := range p.leftEqualityIndices {
if p.leftEqualityIndices[i] == leftColIdx && p.rightEqualityIndices[i] == rightColIdx {
// The filter is already there; simply absorb it and say we succeeded.
return true
}
}
// First resolve the comparison function. We can't use the
// ComparisonExpr's memoized comparison directly, because we may
// have swapped the operands above.
fn, found := parser.FindEqualComparisonFunction(lhs.ResolvedType(), rhs.ResolvedType())
if !found {
// This is ... unexpected. This means we have a valid ON
// expression of the form "a = b" but the expression "b = a" is
// invalid. We could simply avoid the optimization but this is
// really a bug in the built-in semantics so we want to complain
// loudly.
panic(fmt.Errorf("predicate %s is valid, but '%T = %T' cannot be type checked", c, lhs, rhs))
}
p.cmpFunctions = append(p.cmpFunctions, fn)
p.leftEqualityIndices = append(p.leftEqualityIndices, leftColIdx)
p.rightEqualityIndices = append(p.rightEqualityIndices, rightColIdx)
p.leftColNames = append(p.leftColNames, parser.Name(left.sourceColumns[leftColIdx].Name))
p.rightColNames = append(p.rightColNames, parser.Name(right.sourceColumns[rightColIdx].Name))
return true
}
// optimizeOnPredicate tries to turn the filter in an onPredicate into
// equality columns in the joinPredicate, which enables faster
// joins. The concatInfos argument, if provided, must be a
// precomputed concatenation of the left and right dataSourceInfos.
func optimizeOnPredicate(
pred *joinPredicate, left, right *dataSourceInfo, concatInfos *dataSourceInfo,
) (*joinPredicate, *dataSourceInfo, error) {
c, ok := pred.filter.(*parser.ComparisonExpr)
if !ok || c.Operator != parser.EQ {
return pred, pred.info, nil
}
lhs, ok := c.Left.(*parser.IndexedVar)
if !ok {
return pred, pred.info, nil
}
rhs, ok := c.Right.(*parser.IndexedVar)
if !ok {
return pred, pred.info, nil
}
sourceBoundary := pred.numMergedEqualityColumns + len(left.sourceColumns)
if (lhs.Idx >= sourceBoundary && rhs.Idx >= sourceBoundary) ||
(lhs.Idx < sourceBoundary && rhs.Idx < sourceBoundary) {
// Both variables are on the same side of the join (e.g. `a JOIN b ON a.x = a.y`).
return pred, pred.info, nil
}
if lhs.Idx > rhs.Idx {
lhs, rhs = rhs, lhs
}
// At this point we have an equality, so we can add it to the list
// of equality columns.
// First resolve the comparison function. We can't use the
// ComparisonExpr's memoized comparison directly, because we may
// have swapped the operands above.
fn, found := parser.FindEqualComparisonFunction(lhs.ResolvedType(), rhs.ResolvedType())
if !found {
// This is ... unexpected. This means we have a valid ON
// expression of the form "a = b" but the expression "b = a" is
// invalid. We could simply avoid the optimization but this is
// really a bug in the built-in semantics so we want to complain
// loudly.
panic(fmt.Errorf("predicate %s is valid, but '%T = %T' cannot be type checked", c, lhs, rhs))
}
pred.cmpFunctions = append(pred.cmpFunctions, fn)
// To do this we must be a bit careful: the expression contains
// IndexedVars, and the column indices at this point will refer to
// the full column set of the joinPredicate, including the
// merged columns.
leftColIdx := lhs.Idx - pred.numMergedEqualityColumns
rightColIdx := rhs.Idx - len(left.sourceColumns) - pred.numMergedEqualityColumns
pred.leftEqualityIndices = append(pred.leftEqualityIndices, leftColIdx)
pred.rightEqualityIndices = append(pred.rightEqualityIndices, rightColIdx)
pred.leftColNames = append(pred.leftColNames, parser.Name(left.sourceColumns[leftColIdx].Name))
pred.rightColNames = append(pred.rightColNames, parser.Name(right.sourceColumns[rightColIdx].Name))
// The filter is optimized away now.
pred.filter = nil
return pred, pred.info, nil
}