Merge pull request #2019 from jackc/fix-encode-driver-valuer-on-pointer

Fix encode driver.Valuer on pointer

README.md

@@ -92,7 +92,7 @@ See the presentation at Golang Estonia, [PGX Top to Bottom](https://www.youtube.
 
 ## Supported Go and PostgreSQL Versions
 
-pgx supports the same versions of Go and PostgreSQL that are supported by their respective teams. For [Go](https://golang.org/doc/devel/release.html#policy) that is the two most recent major releases and for [PostgreSQL](https://www.postgresql.org/support/versioning/) the major releases in the last 5 years. This means pgx supports Go 1.20 and higher and PostgreSQL 12 and higher. pgx also is tested against the latest version of [CockroachDB](https://www.cockroachlabs.com/product/).
+pgx supports the same versions of Go and PostgreSQL that are supported by their respective teams. For [Go](https://golang.org/doc/devel/release.html#policy) that is the two most recent major releases and for [PostgreSQL](https://www.postgresql.org/support/versioning/) the major releases in the last 5 years. This means pgx supports Go 1.21 and higher and PostgreSQL 12 and higher. pgx also is tested against the latest version of [CockroachDB](https://www.cockroachlabs.com/product/).
 
 ## Version Policy
 

conn.go

@@ -10,7 +10,6 @@ import (
 	"strings"
 	"time"
 
-	"github.com/jackc/pgx/v5/internal/anynil"
 	"github.com/jackc/pgx/v5/internal/sanitize"
 	"github.com/jackc/pgx/v5/internal/stmtcache"
 	"github.com/jackc/pgx/v5/pgconn"
@@ -755,7 +754,6 @@ optionLoop:
 	}
 
 	c.eqb.reset()
-	anynil.NormalizeSlice(args)
 	rows := c.getRows(ctx, sql, args)
 
 	var err error

extended_query_builder.go

@@ -1,10 +1,8 @@
 package pgx
 
 import (
-	"database/sql/driver"
 	"fmt"
 
-	"github.com/jackc/pgx/v5/internal/anynil"
 	"github.com/jackc/pgx/v5/pgconn"
 	"github.com/jackc/pgx/v5/pgtype"
 )
@@ -23,10 +21,15 @@ type ExtendedQueryBuilder struct {
 func (eqb *ExtendedQueryBuilder) Build(m *pgtype.Map, sd *pgconn.StatementDescription, args []any) error {
 	eqb.reset()
 
-	anynil.NormalizeSlice(args)
-
 	if sd == nil {
-		return eqb.appendParamsForQueryExecModeExec(m, args)
+		for i := range args {
+			err := eqb.appendParam(m, 0, pgtype.TextFormatCode, args[i])
+			if err != nil {
+				err = fmt.Errorf("failed to encode args[%d]: %w", i, err)
+				return err
+			}
+		}
+		return nil
 	}
 
 	if len(sd.ParamOIDs) != len(args) {
@@ -113,10 +116,6 @@ func (eqb *ExtendedQueryBuilder) reset() {
 }
 
 func (eqb *ExtendedQueryBuilder) encodeExtendedParamValue(m *pgtype.Map, oid uint32, formatCode int16, arg any) ([]byte, error) {
-	if anynil.Is(arg) {
-		return nil, nil
-	}
-
 	if eqb.paramValueBytes == nil {
 		eqb.paramValueBytes = make([]byte, 0, 128)
 	}
@@ -145,74 +144,3 @@ func (eqb *ExtendedQueryBuilder) chooseParameterFormatCode(m *pgtype.Map, oid ui
 
 	return m.FormatCodeForOID(oid)
 }
-
-// appendParamsForQueryExecModeExec appends the args to eqb.
-//
-// Parameters must be encoded in the text format because of differences in type conversion between timestamps and
-// dates. In QueryExecModeExec we don't know what the actual PostgreSQL type is. To determine the type we use the
-// Go type to OID type mapping registered by RegisterDefaultPgType. However, the Go time.Time represents both
-// PostgreSQL timestamp[tz] and date. To use the binary format we would need to also specify what the PostgreSQL
-// type OID is. But that would mean telling PostgreSQL that we have sent a timestamp[tz] when what is needed is a date.
-// This means that the value is converted from text to timestamp[tz] to date. This means it does a time zone conversion
-// before converting it to date. This means that dates can be shifted by one day. In text format without that double
-// type conversion it takes the date directly and ignores time zone (i.e. it works).
-//
-// Given that the whole point of QueryExecModeExec is to operate without having to know the PostgreSQL types there is
-// no way to safely use binary or to specify the parameter OIDs.
-func (eqb *ExtendedQueryBuilder) appendParamsForQueryExecModeExec(m *pgtype.Map, args []any) error {
-	for _, arg := range args {
-		if arg == nil {
-			err := eqb.appendParam(m, 0, TextFormatCode, arg)
-			if err != nil {
-				return err
-			}
-		} else {
-			dt, ok := m.TypeForValue(arg)
-			if !ok {
-				var tv pgtype.TextValuer
-				if tv, ok = arg.(pgtype.TextValuer); ok {
-					t, err := tv.TextValue()
-					if err != nil {
-						return err
-					}
-
-					dt, ok = m.TypeForOID(pgtype.TextOID)
-					if ok {
-						arg = t
-					}
-				}
-			}
-			if !ok {
-				var dv driver.Valuer
-				if dv, ok = arg.(driver.Valuer); ok {
-					v, err := dv.Value()
-					if err != nil {
-						return err
-					}
-					dt, ok = m.TypeForValue(v)
-					if ok {
-						arg = v
-					}
-				}
-			}
-			if !ok {
-				var str fmt.Stringer
-				if str, ok = arg.(fmt.Stringer); ok {
-					dt, ok = m.TypeForOID(pgtype.TextOID)
-					if ok {
-						arg = str.String()
-					}
-				}
-			}
-			if !ok {
-				return &unknownArgumentTypeQueryExecModeExecError{arg: arg}
-			}
-			err := eqb.appendParam(m, dt.OID, TextFormatCode, arg)
-			if err != nil {
-				return err
-			}
-		}
-	}
-
-	return nil
-}

go.mod

@@ -1,6 +1,6 @@
 module github.com/jackc/pgx/v5
 
-go 1.19
+go 1.20
 
 require (
 	github.com/jackc/pgpassfile v1.0.0

pgtype/array_codec.go

@@ -6,7 +6,6 @@ import (
 	"fmt"
 	"reflect"
 
-	"github.com/jackc/pgx/v5/internal/anynil"
 	"github.com/jackc/pgx/v5/internal/pgio"
 )
 
@@ -230,7 +229,7 @@ func (c *ArrayCodec) PlanScan(m *Map, oid uint32, format int16, target any) Scan
 
 	// target / arrayScanner might be a pointer to a nil. If it is create one so we can call ScanIndexType to plan the
 	// scan of the elements.
-	if anynil.Is(target) {
+	if isNil, _ := isNilDriverValuer(target); isNil {
 		arrayScanner = reflect.New(reflect.TypeOf(target).Elem()).Interface().(ArraySetter)
 	}
 

pgtype/doc.go

@@ -139,6 +139,16 @@ Compatibility with database/sql
 pgtype also includes support for custom types implementing the database/sql.Scanner and database/sql/driver.Valuer
 interfaces.
 
+Encoding Typed Nils
+
+pgtype encodes untyped and typed nils (e.g. nil and []byte(nil)) to the SQL NULL value without going through the Codec
+system. This means that Codecs and other encoding logic do not have to handle nil or *T(nil).
+
+However, database/sql compatibility requires Value to be called on T(nil) when T implements driver.Valuer. Therefore,
+driver.Valuer values are only considered NULL when *T(nil) where driver.Valuer is implemented on T not on *T. See
+https://github.com/golang/go/issues/8415 and
+https://github.com/golang/go/commit/0ce1d79a6a771f7449ec493b993ed2a720917870.
+
 Child Records
 
 pgtype's support for arrays and composite records can be used to load records and their children in a single query.  See

pgtype/pgtype.go

@@ -1912,8 +1912,17 @@ func newEncodeError(value any, m *Map, oid uint32, formatCode int16, err error)
 // (nil, nil). The caller of Encode is responsible for writing the correct NULL value or the length of the data
 // written.
 func (m *Map) Encode(oid uint32, formatCode int16, value any, buf []byte) (newBuf []byte, err error) {
-	if value == nil {
-		return nil, nil
+	if isNil, callNilDriverValuer := isNilDriverValuer(value); isNil {
+		if callNilDriverValuer {
+			newBuf, err = (&encodePlanDriverValuer{m: m, oid: oid, formatCode: formatCode}).Encode(value, buf)
+			if err != nil {
+				return nil, newEncodeError(value, m, oid, formatCode, err)
+			}
+
+			return newBuf, nil
+		} else {
+			return nil, nil
+		}
 	}
 
 	plan := m.PlanEncode(oid, formatCode, value)
@@ -1968,3 +1977,55 @@ func (w *sqlScannerWrapper) Scan(src any) error {
 
 	return w.m.Scan(t.OID, TextFormatCode, bufSrc, w.v)
 }
+
+// canBeNil returns true if value can be nil.
+func canBeNil(value any) bool {
+	refVal := reflect.ValueOf(value)
+	kind := refVal.Kind()
+	switch kind {
+	case reflect.Chan, reflect.Func, reflect.Map, reflect.Ptr, reflect.UnsafePointer, reflect.Interface, reflect.Slice:
+		return true
+	default:
+		return false
+	}
+}
+
+// valuerReflectType is a reflect.Type for driver.Valuer. It has confusing syntax because reflect.TypeOf returns nil
+// when it's argument is a nil interface value. So we use a pointer to the interface and call Elem to get the actual
+// type. Yuck.
+//
+// This can be simplified in Go 1.22 with reflect.TypeFor.
+//
+// var valuerReflectType = reflect.TypeFor[driver.Valuer]()
+var valuerReflectType = reflect.TypeOf((*driver.Valuer)(nil)).Elem()
+
+// isNilDriverValuer returns true if value is any type of nil unless it implements driver.Valuer. *T is not considered to implement
+// driver.Valuer if it is only implemented by T.
+func isNilDriverValuer(value any) (isNil bool, callNilDriverValuer bool) {
+	if value == nil {
+		return true, false
+	}
+
+	refVal := reflect.ValueOf(value)
+	kind := refVal.Kind()
+	switch kind {
+	case reflect.Chan, reflect.Func, reflect.Map, reflect.Ptr, reflect.UnsafePointer, reflect.Interface, reflect.Slice:
+		if !refVal.IsNil() {
+			return false, false
+		}
+
+		if _, ok := value.(driver.Valuer); ok {
+			if kind == reflect.Ptr {
+				// The type assertion will succeed if driver.Valuer is implemented on T or *T. Check if it is implemented on *T
+				// by checking if it is not implemented on *T.
+				return true, !refVal.Type().Elem().Implements(valuerReflectType)
+			} else {
+				return true, true
+			}
+		}
+
+		return true, false
+	default:
+		return false, false
+	}
+}