planner: fix the issue that cannot find column if using question mark…

…er in group-by-clause (pingcap#54205) (pingcap#55979)

close pingcap#53872

planner/core/logical_plan_builder.go

@@ -3088,7 +3088,7 @@ func (g *gbyResolver) Enter(inNode ast.Node) (ast.Node, bool) {
 		return inNode, true
 	case *driver.ParamMarkerExpr:
 		g.isParam = true
-		if g.exprDepth == 1 {
+		if g.exprDepth == 1 && !n.UseAsValueInGbyByClause {
 			_, isNull, isExpectedType := getUintFromNode(g.ctx, n, false)
 			// For constant uint expression in top level, it should be treated as position expression.
 			if !isNull && isExpectedType {
@@ -3126,6 +3126,9 @@ func (g *gbyResolver) Leave(inNode ast.Node) (ast.Node, bool) {
 				} else if ast.HasWindowFlag(ret) {
 					err = ErrIllegalReference.GenWithStackByArgs(v.Name.OrigColName(), "reference to window function")
 				} else {
+					if isParam, ok := ret.(*driver.ParamMarkerExpr); ok {
+						isParam.UseAsValueInGbyByClause = true
+					}
 					return ret, true
 				}
 			}

planner/core/plan_cache_test.go

@@ -1132,6 +1132,18 @@ func TestIssue41626(t *testing.T) {
 	tk.MustQuery(`show warnings`).Check(testkit.Rows("Warning 1105 skip prepared plan-cache: '12' may be converted to INT"))
 }
 
+func TestIssue53872(t *testing.T) {
+	store := testkit.CreateMockStore(t)
+	tk := testkit.NewTestKit(t, store)
+
+	tk.MustExec(`use test`)
+	tk.MustExec(`create table test(id int, col int)`)
+	tk.MustExec(`prepare stmt from "select id, ? as col1 from test where col=? group by id,col1"`)
+	tk.MustExec(`set @a=100, @b=100`)
+	tk.MustQuery(`execute stmt using @a,@b`).Check(testkit.Rows()) // no error
+	tk.MustQuery(`execute stmt using @a,@b`).Check(testkit.Rows())
+}
+
 func TestIssue38269(t *testing.T) {
 	store := testkit.CreateMockStore(t)
 	tk := testkit.NewTestKit(t, store)

types/parser_driver/value_expr.go

@@ -240,6 +240,14 @@ type ParamMarkerExpr struct {
 	Offset    int
 	Order     int
 	InExecute bool
+
+	// For "select ? as c from t group by c", the optimizer replaces the `c` in the by-clause to `group by ?`,
+	// but this conversion conflicts with the original semantic. The original `group by c` means grouping by the column `c`,
+	// while the converted `group by ?` means grouping by the `?-th` column in the select-list, for example, `group by 3` means
+	// grouping the result by the 3rd column.
+	// Use this flag to let the optimizer know whether `group by ?` is converted from this case and if it is, use this
+	// marker as normal value instead of column index in the by-clause.
+	UseAsValueInGbyByClause bool
 }
 
 // Restore implements Node interface.