Merge pull request #107 from goccy/feature/refactor-encoder

Refactor the encoder to fix a bug where streaming encoders can't reuse buffer

.codecov.yml

@@ -0,0 +1,28 @@
+codecov:
+  require_ci_to_pass: yes
+
+coverage:
+  precision: 2
+  round: down
+  range: "70...100"
+
+  status:
+    project:
+      default:
+        target: 70%
+        threshold: 2%
+    patch: off
+    changes: no
+
+parsers:
+  gcov:
+    branch_detection:
+      conditional: yes
+      loop: yes
+      method: no
+      macro: no
+
+comment:
+  layout: "header,diff"
+  behavior: default
+  require_changes: no

benchmarks/encode_test.go

@@ -1,6 +1,7 @@
 package benchmark
 
 import (
+	"bytes"
 	"encoding/json"
 	"testing"
 
@@ -559,6 +560,51 @@ func Benchmark_Encode_Interface_GoJson(b *testing.B) {
 }
 
 func Benchmark_Encode_Bool_EncodingJson(b *testing.B) {
+	b.ReportAllocs()
+	var buf bytes.Buffer
+	enc := json.NewEncoder(&buf)
+	for i := 0; i < b.N; i++ {
+		if err := enc.Encode(true); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_Bool_JsonIter(b *testing.B) {
+	var json = jsoniter.ConfigCompatibleWithStandardLibrary
+	b.ReportAllocs()
+	var buf bytes.Buffer
+	enc := json.NewEncoder(&buf)
+	for i := 0; i < b.N; i++ {
+		if err := enc.Encode(true); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_Bool_SegmentioJson(b *testing.B) {
+	b.ReportAllocs()
+	var buf bytes.Buffer
+	enc := segmentiojson.NewEncoder(&buf)
+	for i := 0; i < b.N; i++ {
+		if err := enc.Encode(true); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_Bool_GoJson(b *testing.B) {
+	b.ReportAllocs()
+	var buf bytes.Buffer
+	enc := gojson.NewEncoder(&buf)
+	for i := 0; i < b.N; i++ {
+		if err := enc.Encode(true); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Marshal_Bool_EncodingJson(b *testing.B) {
 	b.ReportAllocs()
 	for i := 0; i < b.N; i++ {
 		if _, err := json.Marshal(true); err != nil {
@@ -567,7 +613,7 @@ func Benchmark_Encode_Bool_EncodingJson(b *testing.B) {
 	}
 }
 
-func Benchmark_Encode_Bool_JsonIter(b *testing.B) {
+func Benchmark_Marshal_Bool_JsonIter(b *testing.B) {
 	var json = jsoniter.ConfigCompatibleWithStandardLibrary
 	b.ReportAllocs()
 	for i := 0; i < b.N; i++ {
@@ -577,7 +623,7 @@ func Benchmark_Encode_Bool_JsonIter(b *testing.B) {
 	}
 }
 
-func Benchmark_Encode_Bool_Jettison(b *testing.B) {
+func Benchmark_Marshal_Bool_Jettison(b *testing.B) {
 	b.ReportAllocs()
 	for i := 0; i < b.N; i++ {
 		if _, err := jettison.Marshal(true); err != nil {
@@ -586,7 +632,7 @@ func Benchmark_Encode_Bool_Jettison(b *testing.B) {
 	}
 }
 
-func Benchmark_Encode_Bool_SegmentioJson(b *testing.B) {
+func Benchmark_Marshal_Bool_SegmentioJson(b *testing.B) {
 	b.ReportAllocs()
 	for i := 0; i < b.N; i++ {
 		if _, err := segmentiojson.Marshal(true); err != nil {
@@ -595,7 +641,7 @@ func Benchmark_Encode_Bool_SegmentioJson(b *testing.B) {
 	}
 }
 
-func Benchmark_Encode_Bool_GoJson(b *testing.B) {
+func Benchmark_Marshal_Bool_GoJson(b *testing.B) {
 	b.ReportAllocs()
 	for i := 0; i < b.N; i++ {
 		if _, err := gojson.Marshal(true); err != nil {

encode.go

@@ -2,154 +2,58 @@ package json
 
 import (
 	"bytes"
-	"encoding"
 	"encoding/base64"
-	"fmt"
 	"io"
 	"math"
-	"reflect"
 	"strconv"
 	"sync"
-	"sync/atomic"
 	"unsafe"
 )
 
 // An Encoder writes JSON values to an output stream.
 type Encoder struct {
 	w                 io.Writer
-	ctx               *encodeRuntimeContext
-	ptr               unsafe.Pointer
-	buf               []byte
 	enabledIndent     bool
 	enabledHTMLEscape bool
-	unorderedMap      bool
+	prefix            string
-	baseIndent        int
+	indentStr         string
-	prefix            []byte
-	indentStr         []byte
-}
-
-type compiledCode struct {
-	code    *opcode
-	linked  bool // whether recursive code already have linked
-	curLen  uintptr
-	nextLen uintptr
 }
 
 const (
 	bufSize = 1024
-	maxAcceptableTypeAddrRange = 1024 * 1024 * 2 // 2 Mib
 )
 
+type EncodeOption int
+
 const (
-	opCodeEscapedType = iota
+	EncodeOptionHTMLEscape EncodeOption = 1 << iota
-	opCodeEscapedIndentType
+	EncodeOptionIndent
-	opCodeNoEscapeType
+	EncodeOptionUnorderedMap
-	opCodeNoEscapeIndentType
 )
 
-type opcodeSet struct {
-	code       *opcode
-	codeLength int
-}
-
-func loadOpcodeMap() map[uintptr]*opcodeSet {
-	p := atomic.LoadPointer(&cachedOpcode)
-	return *(*map[uintptr]*opcodeSet)(unsafe.Pointer(&p))
-}
-
-func storeOpcodeSet(typ uintptr, set *opcodeSet, m map[uintptr]*opcodeSet) {
-	newOpcodeMap := make(map[uintptr]*opcodeSet, len(m)+1)
-	newOpcodeMap[typ] = set
-
-	for k, v := range m {
-		newOpcodeMap[k] = v
-	}
-
-	atomic.StorePointer(&cachedOpcode, *(*unsafe.Pointer)(unsafe.Pointer(&newOpcodeMap)))
-}
-
 var (
-	encPool          sync.Pool
+	encRuntimeContextPool = sync.Pool{
-	codePool         sync.Pool
-	cachedOpcode     unsafe.Pointer // map[uintptr]*opcodeSet
-	marshalJSONType  reflect.Type
-	marshalTextType  reflect.Type
-	baseTypeAddr     uintptr
-	cachedOpcodeSets []*opcodeSet
-)
-
-//go:linkname typelinks reflect.typelinks
-func typelinks() ([]unsafe.Pointer, [][]int32)
-
-//go:linkname rtypeOff reflect.rtypeOff
-func rtypeOff(unsafe.Pointer, int32) unsafe.Pointer
-
-func setupOpcodeSets() error {
-	sections, offsets := typelinks()
-	if len(sections) != 1 {
-		return fmt.Errorf("failed to get sections")
-	}
-	if len(offsets) != 1 {
-		return fmt.Errorf("failed to get offsets")
-	}
-	section := sections[0]
-	offset := offsets[0]
-	var (
-		min uintptr = uintptr(^uint(0))
-		max uintptr = 0
-	)
-	for i := 0; i < len(offset); i++ {
-		addr := uintptr(rtypeOff(section, offset[i]))
-		if min > addr {
-			min = addr
-		}
-		if max < addr {
-			max = addr
-		}
-	}
-	addrRange := uintptr(max) - uintptr(min)
-	if addrRange == 0 {
-		return fmt.Errorf("failed to get address range of types")
-	}
-	if addrRange > maxAcceptableTypeAddrRange {
-		return fmt.Errorf("too big address range %d", addrRange)
-	}
-	cachedOpcodeSets = make([]*opcodeSet, addrRange)
-	baseTypeAddr = min
-	return nil
-}
-
-func init() {
-	encPool = sync.Pool{
 		New: func() interface{} {
-			return &Encoder{
+			return &encodeRuntimeContext{
-				ctx: &encodeRuntimeContext{
+				buf:      make([]byte, 0, bufSize),
 				ptrs:     make([]uintptr, 128),
 				keepRefs: make([]unsafe.Pointer, 0, 8),
-				},
-				buf: make([]byte, 0, bufSize),
 			}
 		},
 	}
-	marshalJSONType = reflect.TypeOf((*Marshaler)(nil)).Elem()
+)
-	marshalTextType = reflect.TypeOf((*encoding.TextMarshaler)(nil)).Elem()
+
-	if err := setupOpcodeSets(); err != nil {
+func takeEncodeRuntimeContext() *encodeRuntimeContext {
-		// fallback to slow path
+	return encRuntimeContextPool.Get().(*encodeRuntimeContext)
-	}
+}
+
+func releaseEncodeRuntimeContext(ctx *encodeRuntimeContext) {
+	encRuntimeContextPool.Put(ctx)
 }
 
 // NewEncoder returns a new encoder that writes to w.
 func NewEncoder(w io.Writer) *Encoder {
-	enc := encPool.Get().(*Encoder)
+	return &Encoder{w: w, enabledHTMLEscape: true}
-	enc.w = w
-	enc.reset()
-	return enc
-}
-
-func newEncoder() *Encoder {
-	enc := encPool.Get().(*Encoder)
-	enc.reset()
-	return enc
 }
 
 // Encode writes the JSON encoding of v to the stream, followed by a newline character.
@@ -160,15 +64,32 @@ func (e *Encoder) Encode(v interface{}) error {
 }
 
 // EncodeWithOption call Encode with EncodeOption.
-func (e *Encoder) EncodeWithOption(v interface{}, opts ...EncodeOption) error {
+func (e *Encoder) EncodeWithOption(v interface{}, optFuncs ...EncodeOptionFunc) error {
-	for _, opt := range opts {
+	ctx := takeEncodeRuntimeContext()
-		if err := opt(e); err != nil {
+
+	err := e.encodeWithOption(ctx, v, optFuncs...)
+
+	releaseEncodeRuntimeContext(ctx)
 	return err
+}
+
+func (e *Encoder) encodeWithOption(ctx *encodeRuntimeContext, v interface{}, optFuncs ...EncodeOptionFunc) error {
+	var opt EncodeOption
+	if e.enabledHTMLEscape {
+		opt |= EncodeOptionHTMLEscape
 	}
+	for _, optFunc := range optFuncs {
+		opt = optFunc(opt)
+	}
+	var (
+		buf []byte
+		err error
+	)
+	if e.enabledIndent {
+		buf, err = encodeIndent(ctx, v, e.prefix, e.indentStr, opt)
+	} else {
+		buf, err = encode(ctx, v, opt)
 	}
-	header := (*interfaceHeader)(unsafe.Pointer(&v))
-	e.ptr = header.ptr
-	buf, err := e.encode(header, v == nil)
 	if err != nil {
 		return err
 	}
@@ -181,7 +102,6 @@ func (e *Encoder) EncodeWithOption(v interface{}, opts ...EncodeOption) error {
 	if _, err := e.w.Write(buf); err != nil {
 		return err
 	}
-	e.buf = buf[:0]
 	return nil
 }
 
@@ -200,140 +120,171 @@ func (e *Encoder) SetIndent(prefix, indent string) {
 		e.enabledIndent = false
 		return
 	}
-	e.prefix = []byte(prefix)
+	e.prefix = prefix
-	e.indentStr = []byte(indent)
+	e.indentStr = indent
 	e.enabledIndent = true
 }
 
-func (e *Encoder) release() {
+func marshal(v interface{}, opt EncodeOption) ([]byte, error) {
-	e.w = nil
+	ctx := takeEncodeRuntimeContext()
-	encPool.Put(e)
-}
 
-func (e *Encoder) reset() {
+	buf, err := encode(ctx, v, EncodeOptionHTMLEscape)
-	e.baseIndent = 0
-	e.enabledHTMLEscape = true
-	e.enabledIndent = false
-	e.unorderedMap = false
-}
-
-func (e *Encoder) encodeForMarshal(header *interfaceHeader, isNil bool) ([]byte, error) {
-	buf, err := e.encode(header, isNil)
 	if err != nil {
+		releaseEncodeRuntimeContext(ctx)
 		return nil, err
 	}
 
-	e.buf = buf
-
-	if e.enabledIndent {
-		// this line's description is the below.
-		buf = buf[:len(buf)-2]
-
-		copied := make([]byte, len(buf))
-		copy(copied, buf)
-		return copied, nil
-	}
-
 	// this line exists to escape call of `runtime.makeslicecopy` .
 	// if use `make([]byte, len(buf)-1)` and `copy(copied, buf)`,
 	// dst buffer size and src buffer size are differrent.
 	// in this case, compiler uses `runtime.makeslicecopy`, but it is slow.
 	buf = buf[:len(buf)-1]
-
 	copied := make([]byte, len(buf))
 	copy(copied, buf)
+
+	releaseEncodeRuntimeContext(ctx)
 	return copied, nil
 }
 
-func (e *Encoder) encode(header *interfaceHeader, isNil bool) ([]byte, error) {
+func marshalNoEscape(v interface{}, opt EncodeOption) ([]byte, error) {
-	b := e.buf[:0]
+	ctx := takeEncodeRuntimeContext()
-	if isNil {
+
-		b = encodeNull(b)
+	buf, err := encodeNoEscape(ctx, v, EncodeOptionHTMLEscape)
-		if e.enabledIndent {
+	if err != nil {
-			b = encodeIndentComma(b)
+		releaseEncodeRuntimeContext(ctx)
-		} else {
+		return nil, err
-			b = encodeComma(b)
 	}
+
+	// this line exists to escape call of `runtime.makeslicecopy` .
+	// if use `make([]byte, len(buf)-1)` and `copy(copied, buf)`,
+	// dst buffer size and src buffer size are differrent.
+	// in this case, compiler uses `runtime.makeslicecopy`, but it is slow.
+	buf = buf[:len(buf)-1]
+	copied := make([]byte, len(buf))
+	copy(copied, buf)
+
+	releaseEncodeRuntimeContext(ctx)
+	return copied, nil
+}
+
+func marshalIndent(v interface{}, prefix, indent string, opt EncodeOption) ([]byte, error) {
+	ctx := takeEncodeRuntimeContext()
+
+	buf, err := encodeIndent(ctx, v, prefix, indent, EncodeOptionHTMLEscape)
+	if err != nil {
+		releaseEncodeRuntimeContext(ctx)
+		return nil, err
+	}
+
+	buf = buf[:len(buf)-2]
+	copied := make([]byte, len(buf))
+	copy(copied, buf)
+
+	releaseEncodeRuntimeContext(ctx)
+	return copied, nil
+}
+
+func encode(ctx *encodeRuntimeContext, v interface{}, opt EncodeOption) ([]byte, error) {
+	b := ctx.buf[:0]
+	if v == nil {
+		b = encodeNull(b)
+		b = encodeComma(b)
 		return b, nil
 	}
+	header := (*interfaceHeader)(unsafe.Pointer(&v))
 	typ := header.typ
 
 	typeptr := uintptr(unsafe.Pointer(typ))
-	codeSet, err := e.compileToGetCodeSet(typeptr)
+	codeSet, err := encodeCompileToGetCodeSet(typeptr)
 	if err != nil {
 		return nil, err
 	}
 
-	ctx := e.ctx
 	p := uintptr(header.ptr)
 	ctx.init(p, codeSet.codeLength)
-	if e.enabledIndent {
+	buf, err := encodeRunCode(ctx, b, codeSet, opt)
-		if e.enabledHTMLEscape {
-			return e.runEscapedIndent(ctx, b, codeSet)
-		} else {
-			return e.runIndent(ctx, b, codeSet)
-		}
-	}
-	if e.enabledHTMLEscape {
-		return e.runEscaped(ctx, b, codeSet)
-	}
-	return e.run(ctx, b, codeSet)
-}
 
-func (e *Encoder) compileToGetCodeSet(typeptr uintptr) (*opcodeSet, error) {
+	ctx.keepRefs = append(ctx.keepRefs, header.ptr)
-	if cachedOpcodeSets == nil {
-		return e.compileToGetCodeSetSlowPath(typeptr)
-	}
-	if codeSet := cachedOpcodeSets[typeptr-baseTypeAddr]; codeSet != nil {
-		return codeSet, nil
-	}
 
-	// noescape trick for header.typ ( reflect.*rtype )
-	copiedType := *(**rtype)(unsafe.Pointer(&typeptr))
-
-	code, err := e.compileHead(&encodeCompileContext{
-		typ:                      copiedType,
-		root:                     true,
-		structTypeToCompiledCode: map[uintptr]*compiledCode{},
-	})
 	if err != nil {
 		return nil, err
 	}
-	code = copyOpcode(code)
+
-	codeLength := code.totalLength()
+	ctx.buf = buf
-	codeSet := &opcodeSet{
+	return buf, nil
-		code:       code,
-		codeLength: codeLength,
-	}
-	cachedOpcodeSets[int(typeptr-baseTypeAddr)] = codeSet
-	return codeSet, nil
 }
 
-func (e *Encoder) compileToGetCodeSetSlowPath(typeptr uintptr) (*opcodeSet, error) {
+func encodeNoEscape(ctx *encodeRuntimeContext, v interface{}, opt EncodeOption) ([]byte, error) {
-	opcodeMap := loadOpcodeMap()
+	b := ctx.buf[:0]
-	if codeSet, exists := opcodeMap[typeptr]; exists {
+	if v == nil {
-		return codeSet, nil
+		b = encodeNull(b)
+		b = encodeComma(b)
+		return b, nil
 	}
+	header := (*interfaceHeader)(unsafe.Pointer(&v))
+	typ := header.typ
 
-	// noescape trick for header.typ ( reflect.*rtype )
+	typeptr := uintptr(unsafe.Pointer(typ))
-	copiedType := *(**rtype)(unsafe.Pointer(&typeptr))
+	codeSet, err := encodeCompileToGetCodeSet(typeptr)
-
-	code, err := e.compileHead(&encodeCompileContext{
-		typ:                      copiedType,
-		root:                     true,
-		structTypeToCompiledCode: map[uintptr]*compiledCode{},
-	})
 	if err != nil {
 		return nil, err
 	}
-	code = copyOpcode(code)
+
-	codeLength := code.totalLength()
+	p := uintptr(header.ptr)
-	codeSet := &opcodeSet{
+	ctx.init(p, codeSet.codeLength)
-		code:       code,
+	buf, err := encodeRunCode(ctx, b, codeSet, opt)
-		codeLength: codeLength,
+	if err != nil {
+		return nil, err
 	}
-	storeOpcodeSet(typeptr, codeSet, opcodeMap)
+
-	return codeSet, nil
+	ctx.buf = buf
+	return buf, nil
+}
+
+func encodeIndent(ctx *encodeRuntimeContext, v interface{}, prefix, indent string, opt EncodeOption) ([]byte, error) {
+	b := ctx.buf[:0]
+	if v == nil {
+		b = encodeNull(b)
+		b = encodeIndentComma(b)
+		return b, nil
+	}
+	header := (*interfaceHeader)(unsafe.Pointer(&v))
+	typ := header.typ
+
+	typeptr := uintptr(unsafe.Pointer(typ))
+	codeSet, err := encodeCompileToGetCodeSet(typeptr)
+	if err != nil {
+		return nil, err
+	}
+
+	p := uintptr(header.ptr)
+	ctx.init(p, codeSet.codeLength)
+	buf, err := encodeRunIndentCode(ctx, b, codeSet, prefix, indent, opt)
+
+	ctx.keepRefs = append(ctx.keepRefs, header.ptr)
+
+	if err != nil {
+		return nil, err
+	}
+
+	ctx.buf = buf
+	return buf, nil
+}
+
+func encodeRunCode(ctx *encodeRuntimeContext, b []byte, codeSet *opcodeSet, opt EncodeOption) ([]byte, error) {
+	if (opt & EncodeOptionHTMLEscape) != 0 {
+		return encodeRunEscaped(ctx, b, codeSet, opt)
+	}
+	return encodeRun(ctx, b, codeSet, opt)
+}
+
+func encodeRunIndentCode(ctx *encodeRuntimeContext, b []byte, codeSet *opcodeSet, prefix, indent string, opt EncodeOption) ([]byte, error) {
+	ctx.prefix = []byte(prefix)
+	ctx.indentStr = []byte(indent)
+	if (opt & EncodeOptionHTMLEscape) != 0 {
+		return encodeRunEscapedIndent(ctx, b, codeSet, opt)
+	}
+	return encodeRunIndent(ctx, b, codeSet, opt)
 }
 
 func encodeFloat32(b []byte, v float32) []byte {
@@ -389,10 +340,10 @@ func appendStructEnd(b []byte) []byte {
 	return append(b, '}', ',')
 }
 
-func (e *Encoder) appendStructEndIndent(b []byte, indent int) []byte {
+func appendStructEndIndent(ctx *encodeRuntimeContext, b []byte, indent int) []byte {
 	b = append(b, '\n')
-	b = append(b, e.prefix...)
+	b = append(b, ctx.prefix...)
-	b = append(b, bytes.Repeat(e.indentStr, e.baseIndent+indent)...)
+	b = append(b, bytes.Repeat(ctx.indentStr, ctx.baseIndent+indent)...)
 	return append(b, '}', ',', '\n')
 }
 
@@ -411,7 +362,7 @@ func encodeByteSlice(b []byte, src []byte) []byte {
 	return append(append(b, buf...), '"')
 }
 
-func (e *Encoder) encodeIndent(b []byte, indent int) []byte {
+func appendIndent(ctx *encodeRuntimeContext, b []byte, indent int) []byte {
-	b = append(b, e.prefix...)
+	b = append(b, ctx.prefix...)
-	return append(b, bytes.Repeat(e.indentStr, e.baseIndent+indent)...)
+	return append(b, bytes.Repeat(ctx.indentStr, ctx.baseIndent+indent)...)
 }

encode_compile.go

@@ -1,23 +1,178 @@
 package json
 
 import (
+	"encoding"
 	"fmt"
 	"reflect"
 	"strings"
+	"sync/atomic"
 	"unsafe"
 )
 
-func (e *Encoder) compileHead(ctx *encodeCompileContext) (*opcode, error) {
+type compiledCode struct {
+	code    *opcode
+	linked  bool // whether recursive code already have linked
+	curLen  uintptr
+	nextLen uintptr
+}
+
+type opcodeSet struct {
+	code       *opcode
+	codeLength int
+}
+
+var (
+	marshalJSONType  = reflect.TypeOf((*Marshaler)(nil)).Elem()
+	marshalTextType  = reflect.TypeOf((*encoding.TextMarshaler)(nil)).Elem()
+	cachedOpcode     unsafe.Pointer // map[uintptr]*opcodeSet
+	baseTypeAddr     uintptr
+	cachedOpcodeSets []*opcodeSet
+)
+
+const (
+	maxAcceptableTypeAddrRange = 1024 * 1024 * 2 // 2 Mib
+)
+
+//go:linkname typelinks reflect.typelinks
+func typelinks() ([]unsafe.Pointer, [][]int32)
+
+//go:linkname rtypeOff reflect.rtypeOff
+func rtypeOff(unsafe.Pointer, int32) unsafe.Pointer
+
+func setupOpcodeSets() error {
+	sections, offsets := typelinks()
+	if len(sections) != 1 {
+		return fmt.Errorf("failed to get sections")
+	}
+	if len(offsets) != 1 {
+		return fmt.Errorf("failed to get offsets")
+	}
+	section := sections[0]
+	offset := offsets[0]
+	var (
+		min uintptr = uintptr(^uint(0))
+		max uintptr = 0
+	)
+	for i := 0; i < len(offset); i++ {
+		typ := (*rtype)(rtypeOff(section, offset[i]))
+		addr := uintptr(unsafe.Pointer(typ))
+		if min > addr {
+			min = addr
+		}
+		if max < addr {
+			max = addr
+		}
+		if typ.Kind() == reflect.Ptr {
+			addr = uintptr(unsafe.Pointer(typ.Elem()))
+			if min > addr {
+				min = addr
+			}
+			if max < addr {
+				max = addr
+			}
+		}
+	}
+	addrRange := uintptr(max) - uintptr(min)
+	if addrRange == 0 {
+		return fmt.Errorf("failed to get address range of types")
+	}
+	if addrRange > maxAcceptableTypeAddrRange {
+		return fmt.Errorf("too big address range %d", addrRange)
+	}
+	cachedOpcodeSets = make([]*opcodeSet, addrRange)
+	baseTypeAddr = min
+	return nil
+}
+
+func init() {
+	if err := setupOpcodeSets(); err != nil {
+		// fallback to slow path
+	}
+}
+
+func encodeCompileToGetCodeSet(typeptr uintptr) (*opcodeSet, error) {
+	if cachedOpcodeSets == nil {
+		return encodeCompileToGetCodeSetSlowPath(typeptr)
+	}
+	if codeSet := cachedOpcodeSets[typeptr-baseTypeAddr]; codeSet != nil {
+		return codeSet, nil
+	}
+
+	// noescape trick for header.typ ( reflect.*rtype )
+	copiedType := *(**rtype)(unsafe.Pointer(&typeptr))
+
+	code, err := encodeCompileHead(&encodeCompileContext{
+		typ:                      copiedType,
+		root:                     true,
+		structTypeToCompiledCode: map[uintptr]*compiledCode{},
+	})
+	if err != nil {
+		return nil, err
+	}
+	code = copyOpcode(code)
+	codeLength := code.totalLength()
+	codeSet := &opcodeSet{
+		code:       code,
+		codeLength: codeLength,
+	}
+	cachedOpcodeSets[int(typeptr-baseTypeAddr)] = codeSet
+	return codeSet, nil
+}
+
+func encodeCompileToGetCodeSetSlowPath(typeptr uintptr) (*opcodeSet, error) {
+	opcodeMap := loadOpcodeMap()
+	if codeSet, exists := opcodeMap[typeptr]; exists {
+		return codeSet, nil
+	}
+
+	// noescape trick for header.typ ( reflect.*rtype )
+	copiedType := *(**rtype)(unsafe.Pointer(&typeptr))
+
+	code, err := encodeCompileHead(&encodeCompileContext{
+		typ:                      copiedType,
+		root:                     true,
+		structTypeToCompiledCode: map[uintptr]*compiledCode{},
+	})
+	if err != nil {
+		return nil, err
+	}
+	code = copyOpcode(code)
+	codeLength := code.totalLength()
+	codeSet := &opcodeSet{
+		code:       code,
+		codeLength: codeLength,
+	}
+	storeOpcodeSet(typeptr, codeSet, opcodeMap)
+	return codeSet, nil
+}
+
+func loadOpcodeMap() map[uintptr]*opcodeSet {
+	p := atomic.LoadPointer(&cachedOpcode)
+	return *(*map[uintptr]*opcodeSet)(unsafe.Pointer(&p))
+}
+
+func storeOpcodeSet(typ uintptr, set *opcodeSet, m map[uintptr]*opcodeSet) {
+	newOpcodeMap := make(map[uintptr]*opcodeSet, len(m)+1)
+	newOpcodeMap[typ] = set
+
+	for k, v := range m {
+		newOpcodeMap[k] = v
+	}
+
+	atomic.StorePointer(&cachedOpcode, *(*unsafe.Pointer)(unsafe.Pointer(&newOpcodeMap)))
+}
+
+func encodeCompileHead(ctx *encodeCompileContext) (*opcode, error) {
 	typ := ctx.typ
 	switch {
 	case typ.Implements(marshalJSONType):
-		return e.compileMarshalJSON(ctx)
+		return encodeCompileMarshalJSON(ctx)
 	case rtype_ptrTo(typ).Implements(marshalJSONType):
-		return e.compileMarshalJSONPtr(ctx)
+		return encodeCompileMarshalJSONPtr(ctx)
 	case typ.Implements(marshalTextType):
-		return e.compileMarshalText(ctx)
+		return encodeCompileMarshalText(ctx)
 	case rtype_ptrTo(typ).Implements(marshalTextType):
-		return e.compileMarshalTextPtr(ctx)
+		return encodeCompileMarshalTextPtr(ctx)
 	}
 	isPtr := false
 	orgType := typ
@@ -26,32 +181,32 @@ func (e *Encoder) compileHead(ctx *encodeCompileContext) (*opcode, error) {
 		isPtr = true
 	}
 	if typ.Kind() == reflect.Map {
-		return e.compileMap(ctx.withType(typ), isPtr)
+		return encodeCompileMap(ctx.withType(typ), isPtr)
 	} else if typ.Kind() == reflect.Struct {
-		code, err := e.compileStruct(ctx.withType(typ), isPtr)
+		code, err := encodeCompileStruct(ctx.withType(typ), isPtr)
 		if err != nil {
 			return nil, err
 		}
-		e.convertHeadOnlyCode(code, isPtr)
+		encodeConvertHeadOnlyCode(code, isPtr)
-		e.optimizeStructEnd(code)
+		encodeOptimizeStructEnd(code)
-		e.linkRecursiveCode(code)
+		encodeLinkRecursiveCode(code)
 		return code, nil
 	} else if isPtr && typ.Implements(marshalTextType) {
 		typ = orgType
 	} else if isPtr && typ.Implements(marshalJSONType) {
 		typ = orgType
 	}
-	code, err := e.compile(ctx.withType(typ))
+	code, err := encodeCompile(ctx.withType(typ))
 	if err != nil {
 		return nil, err
 	}
-	e.convertHeadOnlyCode(code, isPtr)
+	encodeConvertHeadOnlyCode(code, isPtr)
-	e.optimizeStructEnd(code)
+	encodeOptimizeStructEnd(code)
-	e.linkRecursiveCode(code)
+	encodeLinkRecursiveCode(code)
 	return code, nil
 }
 
-func (e *Encoder) linkRecursiveCode(c *opcode) {
+func encodeLinkRecursiveCode(c *opcode) {
 	for code := c; code.op != opEnd && code.op != opStructFieldRecursiveEnd; {
 		switch code.op {
 		case opStructFieldRecursive,
@@ -82,7 +237,7 @@ func (e *Encoder) linkRecursiveCode(c *opcode) {
 			code.jmp.nextLen = nextTotalLength
 			code.jmp.linked = true
 
-			e.linkRecursiveCode(code.jmp.code)
+			encodeLinkRecursiveCode(code.jmp.code)
 			code = code.next
 			continue
 		}
@@ -95,7 +250,7 @@ func (e *Encoder) linkRecursiveCode(c *opcode) {
 	}
 }
 
-func (e *Encoder) optimizeStructEnd(c *opcode) {
+func encodeOptimizeStructEnd(c *opcode) {
 	for code := c; code.op != opEnd; {
 		if code.op == opStructFieldRecursive {
 			// ignore if exists recursive operation
@@ -136,7 +291,7 @@ func (e *Encoder) optimizeStructEnd(c *opcode) {
 	}
 }
 
-func (e *Encoder) convertHeadOnlyCode(c *opcode, isPtrHead bool) {
+func encodeConvertHeadOnlyCode(c *opcode, isPtrHead bool) {
 	if c.nextField == nil {
 		return
 	}
@@ -145,19 +300,19 @@ func (e *Encoder) convertHeadOnlyCode(c *opcode, isPtrHead bool) {
 	}
 	switch c.op {
 	case opStructFieldHead:
-		e.convertHeadOnlyCode(c.next, false)
+		encodeConvertHeadOnlyCode(c.next, false)
 		if !strings.Contains(c.next.op.String(), "Only") {
 			return
 		}
 		c.op = opStructFieldHeadOnly
 	case opStructFieldHeadOmitEmpty:
-		e.convertHeadOnlyCode(c.next, false)
+		encodeConvertHeadOnlyCode(c.next, false)
 		if !strings.Contains(c.next.op.String(), "Only") {
 			return
 		}
 		c.op = opStructFieldHeadOmitEmptyOnly
 	case opStructFieldHeadStringTag:
-		e.convertHeadOnlyCode(c.next, false)
+		encodeConvertHeadOnlyCode(c.next, false)
 		if !strings.Contains(c.next.op.String(), "Only") {
 			return
 		}
@@ -185,7 +340,7 @@ func (e *Encoder) convertHeadOnlyCode(c *opcode, isPtrHead bool) {
 	}
 }
 
-func (e *Encoder) implementsMarshaler(typ *rtype) bool {
+func encodeImplementsMarshaler(typ *rtype) bool {
 	switch {
 	case typ.Implements(marshalJSONType):
 		return true
@@ -199,115 +354,115 @@ func (e *Encoder) implementsMarshaler(typ *rtype) bool {
 	return false
 }
 
-func (e *Encoder) compile(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompile(ctx *encodeCompileContext) (*opcode, error) {
 	typ := ctx.typ
 	switch {
 	case typ.Implements(marshalJSONType):
-		return e.compileMarshalJSON(ctx)
+		return encodeCompileMarshalJSON(ctx)
 	case rtype_ptrTo(typ).Implements(marshalJSONType):
-		return e.compileMarshalJSONPtr(ctx)
+		return encodeCompileMarshalJSONPtr(ctx)
 	case typ.Implements(marshalTextType):
-		return e.compileMarshalText(ctx)
+		return encodeCompileMarshalText(ctx)
 	case rtype_ptrTo(typ).Implements(marshalTextType):
-		return e.compileMarshalTextPtr(ctx)
+		return encodeCompileMarshalTextPtr(ctx)
 	}
 	switch typ.Kind() {
 	case reflect.Ptr:
-		return e.compilePtr(ctx)
+		return encodeCompilePtr(ctx)
 	case reflect.Slice:
 		elem := typ.Elem()
-		if !e.implementsMarshaler(elem) && elem.Kind() == reflect.Uint8 {
+		if !encodeImplementsMarshaler(elem) && elem.Kind() == reflect.Uint8 {
-			return e.compileBytes(ctx)
+			return encodeCompileBytes(ctx)
 		}
-		return e.compileSlice(ctx)
+		return encodeCompileSlice(ctx)
 	case reflect.Array:
-		return e.compileArray(ctx)
+		return encodeCompileArray(ctx)
 	case reflect.Map:
-		return e.compileMap(ctx, true)
+		return encodeCompileMap(ctx, true)
 	case reflect.Struct:
-		return e.compileStruct(ctx, false)
+		return encodeCompileStruct(ctx, false)
 	case reflect.Interface:
-		return e.compileInterface(ctx)
+		return encodeCompileInterface(ctx)
 	case reflect.Int:
-		return e.compileInt(ctx)
+		return encodeCompileInt(ctx)
 	case reflect.Int8:
-		return e.compileInt8(ctx)
+		return encodeCompileInt8(ctx)
 	case reflect.Int16:
-		return e.compileInt16(ctx)
+		return encodeCompileInt16(ctx)
 	case reflect.Int32:
-		return e.compileInt32(ctx)
+		return encodeCompileInt32(ctx)
 	case reflect.Int64:
-		return e.compileInt64(ctx)
+		return encodeCompileInt64(ctx)
 	case reflect.Uint:
-		return e.compileUint(ctx)
+		return encodeCompileUint(ctx)
 	case reflect.Uint8:
-		return e.compileUint8(ctx)
+		return encodeCompileUint8(ctx)
 	case reflect.Uint16:
-		return e.compileUint16(ctx)
+		return encodeCompileUint16(ctx)
 	case reflect.Uint32:
-		return e.compileUint32(ctx)
+		return encodeCompileUint32(ctx)
 	case reflect.Uint64:
-		return e.compileUint64(ctx)
+		return encodeCompileUint64(ctx)
 	case reflect.Uintptr:
-		return e.compileUint(ctx)
+		return encodeCompileUint(ctx)
 	case reflect.Float32:
-		return e.compileFloat32(ctx)
+		return encodeCompileFloat32(ctx)
 	case reflect.Float64:
-		return e.compileFloat64(ctx)
+		return encodeCompileFloat64(ctx)
 	case reflect.String:
-		return e.compileString(ctx)
+		return encodeCompileString(ctx)
 	case reflect.Bool:
-		return e.compileBool(ctx)
+		return encodeCompileBool(ctx)
 	}
 	return nil, &UnsupportedTypeError{Type: rtype2type(typ)}
 }
 
-func (e *Encoder) compileKey(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileKey(ctx *encodeCompileContext) (*opcode, error) {
 	typ := ctx.typ
 	switch {
 	case rtype_ptrTo(typ).Implements(marshalJSONType):
-		return e.compileMarshalJSONPtr(ctx)
+		return encodeCompileMarshalJSONPtr(ctx)
 	case rtype_ptrTo(typ).Implements(marshalTextType):
-		return e.compileMarshalTextPtr(ctx)
+		return encodeCompileMarshalTextPtr(ctx)
 	}
 	switch typ.Kind() {
 	case reflect.Ptr:
-		return e.compilePtr(ctx)
+		return encodeCompilePtr(ctx)
 	case reflect.Interface:
-		return e.compileInterface(ctx)
+		return encodeCompileInterface(ctx)
 	case reflect.String:
-		return e.compileString(ctx)
+		return encodeCompileString(ctx)
 	case reflect.Int:
-		return e.compileIntString(ctx)
+		return encodeCompileIntString(ctx)
 	case reflect.Int8:
-		return e.compileInt8String(ctx)
+		return encodeCompileInt8String(ctx)
 	case reflect.Int16:
-		return e.compileInt16String(ctx)
+		return encodeCompileInt16String(ctx)
 	case reflect.Int32:
-		return e.compileInt32String(ctx)
+		return encodeCompileInt32String(ctx)
 	case reflect.Int64:
-		return e.compileInt64String(ctx)
+		return encodeCompileInt64String(ctx)
 	case reflect.Uint:
-		return e.compileUintString(ctx)
+		return encodeCompileUintString(ctx)
 	case reflect.Uint8:
-		return e.compileUint8String(ctx)
+		return encodeCompileUint8String(ctx)
 	case reflect.Uint16:
-		return e.compileUint16String(ctx)
+		return encodeCompileUint16String(ctx)
 	case reflect.Uint32:
-		return e.compileUint32String(ctx)
+		return encodeCompileUint32String(ctx)
 	case reflect.Uint64:
-		return e.compileUint64String(ctx)
+		return encodeCompileUint64String(ctx)
 	case reflect.Uintptr:
-		return e.compileUintString(ctx)
+		return encodeCompileUintString(ctx)
 	}
 	return nil, &UnsupportedTypeError{Type: rtype2type(typ)}
 }
 
-func (e *Encoder) compilePtr(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompilePtr(ctx *encodeCompileContext) (*opcode, error) {
 	ptrOpcodeIndex := ctx.opcodeIndex
 	ptrIndex := ctx.ptrIndex
 	ctx.incIndex()
-	code, err := e.compile(ctx.withType(ctx.typ.Elem()))
+	code, err := encodeCompile(ctx.withType(ctx.typ.Elem()))
 	if err != nil {
 		return nil, err
 	}
@@ -324,187 +479,187 @@ func (e *Encoder) compilePtr(ctx *encodeCompileContext) (*opcode, error) {
 	return newOpCodeWithNext(c, opPtr, code), nil
 }
 
-func (e *Encoder) compileMarshalJSON(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileMarshalJSON(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opMarshalJSON)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileMarshalJSONPtr(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileMarshalJSONPtr(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx.withType(rtype_ptrTo(ctx.typ)), opMarshalJSON)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileMarshalText(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileMarshalText(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opMarshalText)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileMarshalTextPtr(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileMarshalTextPtr(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx.withType(rtype_ptrTo(ctx.typ)), opMarshalText)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt8(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt8(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt8)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt16(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt16(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt16)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt32(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt32(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt32)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt64(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt64(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt64)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint8(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint8(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint8)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint16(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint16(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint16)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint32(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint32(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint32)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint64(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint64(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint64)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileIntString(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileIntString(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opIntString)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt8String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt8String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt8String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt16String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt16String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt16String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt32String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt32String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt32String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInt64String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInt64String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opInt64String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUintString(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUintString(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUintString)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint8String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint8String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint8String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint16String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint16String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint16String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint32String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint32String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint32String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileUint64String(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileUint64String(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opUint64String)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileFloat32(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileFloat32(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opFloat32)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileFloat64(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileFloat64(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opFloat64)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileString(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileString(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opString)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileBool(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileBool(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opBool)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileBytes(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileBytes(ctx *encodeCompileContext) (*opcode, error) {
 	code := newOpCode(ctx, opBytes)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileInterface(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileInterface(ctx *encodeCompileContext) (*opcode, error) {
 	code := newInterfaceCode(ctx)
 	ctx.incIndex()
 	return code, nil
 }
 
-func (e *Encoder) compileSlice(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileSlice(ctx *encodeCompileContext) (*opcode, error) {
 	ctx.root = false
 	elem := ctx.typ.Elem()
 	size := elem.Size()
@@ -512,7 +667,7 @@ func (e *Encoder) compileSlice(ctx *encodeCompileContext) (*opcode, error) {
 	header := newSliceHeaderCode(ctx)
 	ctx.incIndex()
 
-	code, err := e.compile(ctx.withType(ctx.typ.Elem()).incIndent())
+	code, err := encodeCompile(ctx.withType(ctx.typ.Elem()).incIndent())
 	if err != nil {
 		return nil, err
 	}
@@ -536,7 +691,7 @@ func (e *Encoder) compileSlice(ctx *encodeCompileContext) (*opcode, error) {
 	return (*opcode)(unsafe.Pointer(header)), nil
 }
 
-func (e *Encoder) compileArray(ctx *encodeCompileContext) (*opcode, error) {
+func encodeCompileArray(ctx *encodeCompileContext) (*opcode, error) {
 	ctx.root = false
 	typ := ctx.typ
 	elem := typ.Elem()
@@ -546,7 +701,7 @@ func (e *Encoder) compileArray(ctx *encodeCompileContext) (*opcode, error) {
 	header := newArrayHeaderCode(ctx, alen)
 	ctx.incIndex()
 
-	code, err := e.compile(ctx.withType(elem).incIndent())
+	code, err := encodeCompile(ctx.withType(elem).incIndent())
 	if err != nil {
 		return nil, err
 	}
@@ -585,7 +740,7 @@ func mapiternext(it unsafe.Pointer)
 //go:noescape
 func maplen(m unsafe.Pointer) int
 
-func (e *Encoder) compileMap(ctx *encodeCompileContext, withLoad bool) (*opcode, error) {
+func encodeCompileMap(ctx *encodeCompileContext, withLoad bool) (*opcode, error) {
 	// header => code => value => code => key => code => value => code => end
 	//                                     ^                       |
 	//                                     |_______________________|
@@ -595,7 +750,7 @@ func (e *Encoder) compileMap(ctx *encodeCompileContext, withLoad bool) (*opcode,
 
 	typ := ctx.typ
 	keyType := ctx.typ.Key()
-	keyCode, err := e.compileKey(ctx.withType(keyType))
+	keyCode, err := encodeCompileKey(ctx.withType(keyType))
 	if err != nil {
 		return nil, err
 	}
@@ -604,7 +759,7 @@ func (e *Encoder) compileMap(ctx *encodeCompileContext, withLoad bool) (*opcode,
 	ctx.incIndex()
 
 	valueType := typ.Elem()
-	valueCode, err := e.compile(ctx.withType(valueType))
+	valueCode, err := encodeCompile(ctx.withType(valueType))
 	if err != nil {
 		return nil, err
 	}
@@ -633,7 +788,7 @@ func (e *Encoder) compileMap(ctx *encodeCompileContext, withLoad bool) (*opcode,
 	return (*opcode)(unsafe.Pointer(header)), nil
 }
 
-func (e *Encoder) typeToHeaderType(ctx *encodeCompileContext, code *opcode) opType {
+func encodeTypeToHeaderType(ctx *encodeCompileContext, code *opcode) opType {
 	switch code.op {
 	case opPtr:
 		ptrNum := 1
@@ -757,7 +912,7 @@ func (e *Encoder) typeToHeaderType(ctx *encodeCompileContext, code *opcode) opTy
 	return opStructFieldHead
 }
 
-func (e *Encoder) typeToFieldType(ctx *encodeCompileContext, code *opcode) opType {
+func encodeTypeToFieldType(ctx *encodeCompileContext, code *opcode) opType {
 	switch code.op {
 	case opPtr:
 		ptrNum := 1
@@ -881,8 +1036,8 @@ func (e *Encoder) typeToFieldType(ctx *encodeCompileContext, code *opcode) opTyp
 	return opStructField
 }
 
-func (e *Encoder) optimizeStructHeader(ctx *encodeCompileContext, code *opcode, tag *structTag) opType {
+func encodeOptimizeStructHeader(ctx *encodeCompileContext, code *opcode, tag *structTag) opType {
-	headType := e.typeToHeaderType(ctx, code)
+	headType := encodeTypeToHeaderType(ctx, code)
 	switch {
 	case tag.isOmitEmpty:
 		headType = headType.headToOmitEmptyHead()
@@ -892,8 +1047,8 @@ func (e *Encoder) optimizeStructHeader(ctx *encodeCompileContext, code *opcode,
 	return headType
 }
 
-func (e *Encoder) optimizeStructField(ctx *encodeCompileContext, code *opcode, tag *structTag) opType {
+func encodeOptimizeStructField(ctx *encodeCompileContext, code *opcode, tag *structTag) opType {
-	fieldType := e.typeToFieldType(ctx, code)
+	fieldType := encodeTypeToFieldType(ctx, code)
 	switch {
 	case tag.isOmitEmpty:
 		fieldType = fieldType.fieldToOmitEmptyField()
@@ -903,24 +1058,24 @@ func (e *Encoder) optimizeStructField(ctx *encodeCompileContext, code *opcode, t
 	return fieldType
 }
 
-func (e *Encoder) recursiveCode(ctx *encodeCompileContext, jmp *compiledCode) *opcode {
+func encodeRecursiveCode(ctx *encodeCompileContext, jmp *compiledCode) *opcode {
 	code := newRecursiveCode(ctx, jmp)
 	ctx.incIndex()
 	return code
 }
 
-func (e *Encoder) compiledCode(ctx *encodeCompileContext) *opcode {
+func encodeCompiledCode(ctx *encodeCompileContext) *opcode {
 	typ := ctx.typ
 	typeptr := uintptr(unsafe.Pointer(typ))
 	if compiledCode, exists := ctx.structTypeToCompiledCode[typeptr]; exists {
-		return e.recursiveCode(ctx, compiledCode)
+		return encodeRecursiveCode(ctx, compiledCode)
 	}
 	return nil
 }
 
-func (e *Encoder) structHeader(ctx *encodeCompileContext, fieldCode *opcode, valueCode *opcode, tag *structTag) *opcode {
+func encodeStructHeader(ctx *encodeCompileContext, fieldCode *opcode, valueCode *opcode, tag *structTag) *opcode {
 	fieldCode.indent--
-	op := e.optimizeStructHeader(ctx, valueCode, tag)
+	op := encodeOptimizeStructHeader(ctx, valueCode, tag)
 	fieldCode.op = op
 	fieldCode.ptrNum = valueCode.ptrNum
 	switch op {
@@ -943,9 +1098,9 @@ func (e *Encoder) structHeader(ctx *encodeCompileContext, fieldCode *opcode, val
 	return (*opcode)(unsafe.Pointer(fieldCode))
 }
 
-func (e *Encoder) structField(ctx *encodeCompileContext, fieldCode *opcode, valueCode *opcode, tag *structTag) *opcode {
+func encodeStructField(ctx *encodeCompileContext, fieldCode *opcode, valueCode *opcode, tag *structTag) *opcode {
 	code := (*opcode)(unsafe.Pointer(fieldCode))
-	op := e.optimizeStructField(ctx, valueCode, tag)
+	op := encodeOptimizeStructField(ctx, valueCode, tag)
 	fieldCode.op = op
 	fieldCode.ptrNum = valueCode.ptrNum
 	switch op {
@@ -968,7 +1123,7 @@ func (e *Encoder) structField(ctx *encodeCompileContext, fieldCode *opcode, valu
 	return code
 }
 
-func (e *Encoder) isNotExistsField(head *opcode) bool {
+func encodeIsNotExistsField(head *opcode) bool {
 	if head == nil {
 		return false
 	}
@@ -990,10 +1145,10 @@ func (e *Encoder) isNotExistsField(head *opcode) bool {
 	if head.next.op.codeType() != codeStructField {
 		return false
 	}
-	return e.isNotExistsField(head.next)
+	return encodeIsNotExistsField(head.next)
 }
 
-func (e *Encoder) optimizeAnonymousFields(head *opcode) {
+func encodeOptimizeAnonymousFields(head *opcode) {
 	code := head
 	var prev *opcode
 	removedFields := map[*opcode]struct{}{}
@@ -1004,13 +1159,13 @@ func (e *Encoder) optimizeAnonymousFields(head *opcode) {
 		if code.op == opStructField {
 			codeType := code.next.op.codeType()
 			if codeType == codeStructField {
-				if e.isNotExistsField(code.next) {
+				if encodeIsNotExistsField(code.next) {
 					code.next = code.nextField
 					diff := code.next.displayIdx - code.displayIdx
 					for i := 0; i < diff; i++ {
 						code.next.decOpcodeIndex()
 					}
-					linkPrevToNextField(code, removedFields)
+					encodeLinkPrevToNextField(code, removedFields)
 					code = prev
 				}
 			}
@@ -1027,7 +1182,7 @@ type structFieldPair struct {
 	linked      bool
 }
 
-func (e *Encoder) anonymousStructFieldPairMap(typ *rtype, tags structTags, named string, valueCode *opcode) map[string][]structFieldPair {
+func encodeAnonymousStructFieldPairMap(typ *rtype, tags structTags, named string, valueCode *opcode) map[string][]structFieldPair {
 	anonymousFields := map[string][]structFieldPair{}
 	f := valueCode
 	var prevAnonymousField *opcode
@@ -1050,7 +1205,7 @@ func (e *Encoder) anonymousStructFieldPairMap(typ *rtype, tags structTags, named
 			for i := 0; i < diff; i++ {
 				f.nextField.decOpcodeIndex()
 			}
-			linkPrevToNextField(f, removedFields)
+			encodeLinkPrevToNextField(f, removedFields)
 		}
 
 		if f.displayKey == "" {
@@ -1069,7 +1224,7 @@ func (e *Encoder) anonymousStructFieldPairMap(typ *rtype, tags structTags, named
 			isTaggedKey: f.isTaggedKey,
 		})
 		if f.next != nil && f.nextField != f.next && f.next.op.codeType() == codeStructField {
-			for k, v := range e.anonymousStructFieldPairMap(typ, tags, named, f.next) {
+			for k, v := range encodeAnonymousStructFieldPairMap(typ, tags, named, f.next) {
 				anonymousFields[k] = append(anonymousFields[k], v...)
 			}
 		}
@@ -1082,7 +1237,7 @@ func (e *Encoder) anonymousStructFieldPairMap(typ *rtype, tags structTags, named
 	return anonymousFields
 }
 
-func (e *Encoder) optimizeConflictAnonymousFields(anonymousFields map[string][]structFieldPair) {
+func encodeOptimizeConflictAnonymousFields(anonymousFields map[string][]structFieldPair) {
 	removedFields := map[*opcode]struct{}{}
 	for _, fieldPairs := range anonymousFields {
 		if len(fieldPairs) == 1 {
@@ -1104,7 +1259,7 @@ func (e *Encoder) optimizeConflictAnonymousFields(anonymousFields map[string][]s
 							fieldPair.curField.nextField.decOpcodeIndex()
 						}
 						removedFields[fieldPair.curField] = struct{}{}
-						linkPrevToNextField(fieldPair.curField, removedFields)
+						encodeLinkPrevToNextField(fieldPair.curField, removedFields)
 					}
 					fieldPair.linked = true
 				}
@@ -1122,7 +1277,7 @@ func (e *Encoder) optimizeConflictAnonymousFields(anonymousFields map[string][]s
 						for i := 0; i < diff; i++ {
 							fieldPair.curField.nextField.decOpcodeIndex()
 						}
-						linkPrevToNextField(fieldPair.curField, removedFields)
+						encodeLinkPrevToNextField(fieldPair.curField, removedFields)
 					}
 					fieldPair.linked = true
 				}
@@ -1135,9 +1290,9 @@ func (e *Encoder) optimizeConflictAnonymousFields(anonymousFields map[string][]s
 	}
 }
 
-func (e *Encoder) compileStruct(ctx *encodeCompileContext, isPtr bool) (*opcode, error) {
+func encodeCompileStruct(ctx *encodeCompileContext, isPtr bool) (*opcode, error) {
 	ctx.root = false
-	if code := e.compiledCode(ctx); code != nil {
+	if code := encodeCompiledCode(ctx); code != nil {
 		return code, nil
 	}
 	typ := ctx.typ
@@ -1179,7 +1334,7 @@ func (e *Encoder) compileStruct(ctx *encodeCompileContext, isPtr bool) (*opcode,
 		fieldOpcodeIndex := ctx.opcodeIndex
 		fieldPtrIndex := ctx.ptrIndex
 		ctx.incIndex()
-		valueCode, err := e.compile(ctx.withType(fieldType))
+		valueCode, err := encodeCompile(ctx.withType(fieldType))
 		if err != nil {
 			return nil, err
 		}
@@ -1195,7 +1350,7 @@ func (e *Encoder) compileStruct(ctx *encodeCompileContext, isPtr bool) (*opcode,
 			if tag.isTaggedKey {
 				tagKey = tag.key
 			}
-			for k, v := range e.anonymousStructFieldPairMap(typ, tags, tagKey, valueCode) {
+			for k, v := range encodeAnonymousStructFieldPairMap(typ, tags, tagKey, valueCode) {
 				anonymousFields[k] = append(anonymousFields[k], v...)
 			}
 		}
@@ -1216,13 +1371,13 @@ func (e *Encoder) compileStruct(ctx *encodeCompileContext, isPtr bool) (*opcode,
 		}
 		if fieldIdx == 0 {
 			fieldCode.headIdx = fieldCode.idx
-			code = e.structHeader(ctx, fieldCode, valueCode, tag)
+			code = encodeStructHeader(ctx, fieldCode, valueCode, tag)
 			head = fieldCode
 			prevField = fieldCode
 		} else {
 			fieldCode.headIdx = head.headIdx
 			code.next = fieldCode
-			code = e.structField(ctx, fieldCode, valueCode, tag)
+			code = encodeStructField(ctx, fieldCode, valueCode, tag)
 			prevField.nextField = fieldCode
 			fieldCode.prevField = prevField
 			prevField = fieldCode
@@ -1265,8 +1420,8 @@ func (e *Encoder) compileStruct(ctx *encodeCompileContext, isPtr bool) (*opcode,
 
 	head.end = structEndCode
 	code.next = structEndCode
-	e.optimizeConflictAnonymousFields(anonymousFields)
+	encodeOptimizeConflictAnonymousFields(anonymousFields)
-	e.optimizeAnonymousFields(head)
+	encodeOptimizeAnonymousFields(head)
 	ret := (*opcode)(unsafe.Pointer(head))
 	compiled.code = ret
 

encode_context.go

@@ -142,9 +142,13 @@ func (c *encodeCompileContext) decPtrIndex() {
 }
 
 type encodeRuntimeContext struct {
+	buf        []byte
 	ptrs       []uintptr
 	keepRefs   []unsafe.Pointer
 	seenPtr    []uintptr
+	baseIndent int
+	prefix     []byte
+	indentStr  []byte
 }
 
 func (c *encodeRuntimeContext) init(p uintptr, codelen int) {
@@ -154,6 +158,7 @@ func (c *encodeRuntimeContext) init(p uintptr, codelen int) {
 	c.ptrs[0] = p
 	c.keepRefs = c.keepRefs[:0]
 	c.seenPtr = c.seenPtr[:0]
+	c.baseIndent = 0
 }
 
 func (c *encodeRuntimeContext) ptr() uintptr {

encode_opcode.go

@@ -327,7 +327,7 @@ func nextField(code *opcode, removedFields map[*opcode]struct{}) *opcode {
 	return code
 }
 
-func linkPrevToNextField(cur *opcode, removedFields map[*opcode]struct{}) {
+func encodeLinkPrevToNextField(cur *opcode, removedFields map[*opcode]struct{}) {
 	prev := prevField(cur.prevField, removedFields)
 	prev.nextField = nextField(cur.nextField, removedFields)
 	code := prev

json.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"errors"
 	"strconv"
-	"unsafe"
 )
 
 // Marshaler is the interface implemented by types that
@@ -160,34 +159,16 @@ func Marshal(v interface{}) ([]byte, error) {
 
 // MarshalNoEscape
 func MarshalNoEscape(v interface{}) ([]byte, error) {
-	enc := newEncoder()
+	return marshalNoEscape(v, EncodeOptionHTMLEscape)
-	header := (*interfaceHeader)(unsafe.Pointer(&v))
-	bytes, err := enc.encodeForMarshal(header, v == nil)
-	if err != nil {
-		enc.release()
-		return nil, err
-	}
-	enc.release()
-	return bytes, nil
 }
 
 // MarshalWithOption returns the JSON encoding of v with EncodeOption.
-func MarshalWithOption(v interface{}, opts ...EncodeOption) ([]byte, error) {
+func MarshalWithOption(v interface{}, optFuncs ...EncodeOptionFunc) ([]byte, error) {
-	enc := newEncoder()
+	opt := EncodeOptionHTMLEscape
-	for _, opt := range opts {
+	for _, optFunc := range optFuncs {
-		if err := opt(enc); err != nil {
+		opt = optFunc(opt)
-			return nil, err
 	}
-	}
+	return marshal(v, opt)
-	header := (*interfaceHeader)(unsafe.Pointer(&v))
-	enc.ptr = header.ptr
-	bytes, err := enc.encodeForMarshal(header, v == nil)
-	if err != nil {
-		enc.release()
-		return nil, err
-	}
-	enc.release()
-	return bytes, nil
 }
 
 // MarshalIndent is like Marshal but applies Indent to format the output.
@@ -198,24 +179,12 @@ func MarshalIndent(v interface{}, prefix, indent string) ([]byte, error) {
 }
 
 // MarshalIndentWithOption is like Marshal but applies Indent to format the output with EncodeOption.
-func MarshalIndentWithOption(v interface{}, prefix, indent string, opts ...EncodeOption) ([]byte, error) {
+func MarshalIndentWithOption(v interface{}, prefix, indent string, optFuncs ...EncodeOptionFunc) ([]byte, error) {
-	var b *bytes.Buffer
+	opt := EncodeOptionHTMLEscape | EncodeOptionIndent
-	enc := NewEncoder(b)
+	for _, optFunc := range optFuncs {
-	for _, opt := range opts {
+		opt = optFunc(opt)
-		if err := opt(enc); err != nil {
-			return nil, err
 	}
-	}
+	return marshalIndent(v, prefix, indent, opt)
-	enc.SetIndent(prefix, indent)
-	header := (*interfaceHeader)(unsafe.Pointer(&v))
-	enc.ptr = header.ptr
-	bytes, err := enc.encodeForMarshal(header, v == nil)
-	if err != nil {
-		enc.release()
-		return nil, err
-	}
-	enc.release()
-	return bytes, nil
 }
 
 // Unmarshal parses the JSON-encoded data and stores the result
@@ -408,12 +377,8 @@ func HTMLEscape(dst *bytes.Buffer, src []byte) {
 	if err := dec.Decode(&v); err != nil {
 		return
 	}
-	enc := NewEncoder(dst)
+	buf, _ := marshal(v, EncodeOptionHTMLEscape)
-	enc.SetEscapeHTML(true)
+	dst.Write(buf)
-	header := (*interfaceHeader)(unsafe.Pointer(&v))
-	enc.ptr = header.ptr
-	enc.buf, _ = enc.encode(header, v == nil)
-	dst.Write(enc.buf[:len(enc.buf)-1]) // remove last ',' character
 }
 
 // Valid reports whether data is a valid JSON encoding.

option.go

@@ -1,10 +1,9 @@
 package json
 
-type EncodeOption func(*Encoder) error
+type EncodeOptionFunc func(EncodeOption) EncodeOption
 
-func UnorderedMap() EncodeOption {
+func UnorderedMap() func(EncodeOption) EncodeOption {
-	return func(e *Encoder) error {
+	return func(opt EncodeOption) EncodeOption {
-		e.unorderedMap = true
+		return opt | EncodeOptionUnorderedMap
-		return nil
 	}
 }