Notice this comment in optbuilder:
|
case *tree.BinaryExpr: |
|
// It's possible for an overload to be selected that expects different |
|
// types than the TypedExpr arguments return: |
|
// |
|
// ARRAY[1, 2] || NULL |
|
// |
|
// This is a tricky case, because the type checker selects []int as the |
|
// type of the right argument, but then types it as unknown. This causes |
|
// issues for the execbuilder, which doesn't have enough information to |
|
// select the right overload. The solution is to wrap any mismatched |
|
// arguments with a CastExpr that preserves the static type. |
|
|
|
left := reType(t.TypedLeft(), t.ResolvedBinOp().LeftType) |
|
right := reType(t.TypedRight(), t.ResolvedBinOp().RightType) |
It sounds like we could avoid these casts entirely if we attached t.ResolvedBinOp().[Left|Right]Type to the possible NullExprs constructed in b.buildScalar below this comment. If the Null was typed, then execbuilder should be able to determine the correct overload.
Jira issue: CRDB-18056