Fix

benchmarks/encode_test.go

@@ -472,3 +472,68 @@ func Benchmark_Encode_LargeStructCached_GoJsonNoEscape(b *testing.B) {
 		}
 	}
 }
+
+func benchMapValue() map[string]interface{} {
+	return map[string]interface{}{
+		"a": 1,
+		"b": 2.1,
+		"c": "hello",
+		"d": struct {
+			V int
+		}{
+			V: 1,
+		},
+		"e": true,
+	}
+}
+
+func Benchmark_Encode_MapInterface_EncodingJson(b *testing.B) {
+	v := benchMapValue()
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		if _, err := json.Marshal(v); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_MapInterface_JsonIter(b *testing.B) {
+	v := benchMapValue()
+	var json = jsoniter.ConfigCompatibleWithStandardLibrary
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		if _, err := json.Marshal(v); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_MapInterface_Jettison(b *testing.B) {
+	v := benchMapValue()
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		if _, err := jettison.Marshal(v); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_MapInterface_SegmentioJson(b *testing.B) {
+	v := benchMapValue()
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		if _, err := segmentiojson.Marshal(v); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func Benchmark_Encode_MapInterface_GoJson(b *testing.B) {
+	v := benchMapValue()
+	b.ReportAllocs()
+	for i := 0; i < b.N; i++ {
+		if _, err := gojson.Marshal(v); err != nil {
+			b.Fatal(err)
+		}
+	}
+}

encode.go

@@ -112,6 +112,7 @@ func (e *Encoder) EncodeWithOption(v interface{}, opts ...EncodeOption) error {
 		}
 	}
 	header := (*interfaceHeader)(unsafe.Pointer(&v))
+	e.ptr = header.ptr
 	buf, err := e.encode(header, v == nil)
 	if err != nil {
 		return err
@@ -192,23 +193,31 @@ func (e *Encoder) encode(header *interfaceHeader, isNil bool) ([]byte, error) {
 	typ := header.typ
 
 	typeptr := uintptr(unsafe.Pointer(typ))
+	codeSet, err := e.compileToGetCodeSet(typeptr)
+	if err != nil {
+		return nil, err
+	}
+
+	ctx := e.ctx
+	p := uintptr(header.ptr)
+	ctx.init(p, codeSet.codeLength)
+	if e.enabledIndent {
+		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) {
 	opcodeMap := loadOpcodeMap()
 	if codeSet, exists := opcodeMap[typeptr]; exists {
-		ctx := e.ctx
-		p := uintptr(header.ptr)
-		ctx.init(p, codeSet.codeLength)
-
-		if e.enabledIndent {
-			if e.enabledHTMLEscape {
-				return e.runEscapedIndent(ctx, b, codeSet.code)
-			} else {
-				return e.runIndent(ctx, b, codeSet.code)
-			}
-		}
-		if e.enabledHTMLEscape {
-			return e.runEscaped(ctx, b, codeSet.code)
-		}
-		return e.run(ctx, b, codeSet.code)
+		return codeSet, nil
 	}
 
 	// noescape trick for header.typ ( reflect.*rtype )
@@ -230,21 +239,7 @@ func (e *Encoder) encode(header *interfaceHeader, isNil bool) ([]byte, error) {
 	}
 
 	storeOpcodeSet(typeptr, codeSet, opcodeMap)
-	p := uintptr(header.ptr)
-	ctx := e.ctx
-	ctx.init(p, codeLength)
-
-	if e.enabledIndent {
-		if e.enabledHTMLEscape {
-			return e.runEscapedIndent(ctx, b, codeSet.code)
-		} else {
-			return e.runIndent(ctx, b, codeSet.code)
-		}
-	}
-	if e.enabledHTMLEscape {
-		return e.runEscaped(ctx, b, codeSet.code)
-	}
-	return e.run(ctx, b, codeSet.code)
+	return codeSet, nil
 }
 
 func encodeFloat32(b []byte, v float32) []byte {

encode_context.go

@@ -1,9 +1,67 @@
 package json
 
 import (
+	"bytes"
+	"sync"
 	"unsafe"
 )
 
+type mapItem struct {
+	key   []byte
+	value []byte
+}
+
+type mapslice struct {
+	items []mapItem
+}
+
+func (m *mapslice) Len() int {
+	return len(m.items)
+}
+
+func (m *mapslice) Less(i, j int) bool {
+	return bytes.Compare(m.items[i].key, m.items[j].key) < 0
+}
+
+func (m *mapslice) Swap(i, j int) {
+	m.items[i], m.items[j] = m.items[j], m.items[i]
+}
+
+type encodeMapContext struct {
+	iter  unsafe.Pointer
+	pos   []int
+	slice *mapslice
+	buf   []byte
+}
+
+var mapContextPool = sync.Pool{
+	New: func() interface{} {
+		return &encodeMapContext{}
+	},
+}
+
+func newMapContext(mapLen int) *encodeMapContext {
+	ctx := mapContextPool.Get().(*encodeMapContext)
+	if ctx.slice == nil {
+		ctx.slice = &mapslice{
+			items: make([]mapItem, 0, mapLen),
+		}
+	}
+	if cap(ctx.pos) < mapLen*2 {
+		ctx.pos = make([]int, 0, mapLen*2)
+		ctx.slice.items = make([]mapItem, 0, mapLen)
+	} else {
+		ctx.pos = ctx.pos[:0]
+		ctx.slice.items = ctx.slice.items[:0]
+	}
+	ctx.buf = ctx.buf[:0]
+	return ctx
+}
+
+func releaseMapContext(c *encodeMapContext) {
+	mapContextPool.Put(c)
+}
+
 type encodeCompileContext struct {
 	typ                      *rtype
 	root                     bool

encode_vm.go

@@ -49,10 +49,11 @@ func errMarshaler(code *opcode, err error) *MarshalerError {
 	}
 }
 
-func (e *Encoder) run(ctx *encodeRuntimeContext, b []byte, code *opcode) ([]byte, error) {
+func (e *Encoder) run(ctx *encodeRuntimeContext, b []byte, codeSet *opcodeSet) ([]byte, error) {
 	recursiveLevel := 0
 	ptrOffset := uintptr(0)
 	ctxptr := ctx.ptr()
+	code := codeSet.code
 
 	for {
 		switch code.op {

encode_vm_escaped.go

@@ -11,11 +11,13 @@ import (
 	"unsafe"
 )
 
-func (e *Encoder) runEscaped(ctx *encodeRuntimeContext, b []byte, code *opcode) ([]byte, error) {
+func (e *Encoder) runEscaped(ctx *encodeRuntimeContext, b []byte, codeSet *opcodeSet) ([]byte, error) {
 	recursiveLevel := 0
 	ptrOffset := uintptr(0)
 	ctxptr := ctx.ptr()
 
+	code := codeSet.code
+
 	for {
 		switch code.op {
 		default:
@@ -167,63 +169,57 @@ func (e *Encoder) runEscaped(ctx *encodeRuntimeContext, b []byte, code *opcode)
 					return nil, errUnsupportedValue(code, ptr)
 				}
 			}
-			ctx.seenPtr = append(ctx.seenPtr, ptr)
-			v := e.ptrToInterface(code, ptr)
-			ctx.keepRefs = append(ctx.keepRefs, unsafe.Pointer(&v))
-			rv := reflect.ValueOf(v)
-			if rv.IsNil() {
+			iface := (*interfaceHeader)(e.ptrToUnsafePtr(ptr))
+			if iface == nil {
 				b = encodeNull(b)
 				b = encodeComma(b)
 				code = code.next
 				break
 			}
-			vv := rv.Interface()
-			header := (*interfaceHeader)(unsafe.Pointer(&vv))
-			if header.typ.Kind() == reflect.Ptr {
-				if rv.Elem().IsNil() {
-					b = encodeNull(b)
-					b = encodeComma(b)
-					code = code.next
-					break
-				}
+			ctx.keepRefs = append(ctx.keepRefs, unsafe.Pointer(iface))
+			if iface.ptr == nil {
+				b = encodeNull(b)
+				b = encodeComma(b)
+				code = code.next
+				break
 			}
-			c, err := e.compileHead(&encodeCompileContext{
-				typ:                      header.typ,
-				root:                     code.root,
-				indent:                   code.indent,
-				structTypeToCompiledCode: map[uintptr]*compiledCode{},
-			})
+			ifaceCodeSet, err := e.compileToGetCodeSet(uintptr(unsafe.Pointer(iface.typ)))
 			if err != nil {
 				return nil, err
 			}
-			beforeLastCode := c.beforeLastCode()
-			lastCode := beforeLastCode.next
-			lastCode.idx = beforeLastCode.idx + uintptrSize
-			totalLength := uintptr(code.totalLength())
-			nextTotalLength := uintptr(c.totalLength())
+
+			totalLength := uintptr(codeSet.codeLength)
+			nextTotalLength := uintptr(ifaceCodeSet.codeLength)
+
 			curlen := uintptr(len(ctx.ptrs))
 			offsetNum := ptrOffset / uintptrSize
-			oldOffset := ptrOffset
 			ptrOffset += totalLength * uintptrSize
 
 			newLen := offsetNum + totalLength + nextTotalLength
 			if curlen < newLen {
 				ctx.ptrs = append(ctx.ptrs, make([]uintptr, newLen-curlen)...)
 			}
-			ctxptr = ctx.ptr() + ptrOffset // assign new ctxptr
 
-			store(ctxptr, 0, uintptr(header.ptr))
-			store(ctxptr, lastCode.idx, oldOffset)
+			newPtrs := ctx.ptrs[ptrOffset/uintptrSize:]
+			newPtrs[0] = uintptr(iface.ptr)
 
-			// link lastCode ( opInterfaceEnd ) => code.next
-			lastCode.op = opInterfaceEnd
-			lastCode.next = code.next
+			oldPtrs := ctx.ptrs
+			ctx.ptrs = newPtrs
+			ctx.seenPtr = append(ctx.seenPtr, uintptr(ptr))
+			bb, err := e.runEscaped(ctx, b, ifaceCodeSet)
+			if err != nil {
+				return nil, err
+			}
 
-			code = c
-			recursiveLevel++
+			ctx.ptrs = oldPtrs
+			ctx.seenPtr = ctx.seenPtr[:len(ctx.seenPtr)-1]
+
+			b = bb
+			code = code.next
 		case opInterfaceEnd:
 			recursiveLevel--
 			// restore ctxptr
+			ctx.seenPtr = ctx.seenPtr[:len(ctx.seenPtr)-1]
 			offset := load(ctxptr, code.idx)
 			ctxptr = ctx.ptr() + offset
 			ptrOffset = offset
@@ -433,6 +429,7 @@ func (e *Encoder) runEscaped(ctx *encodeRuntimeContext, b []byte, code *opcode)
 				ptr := load(ctxptr, code.end.mapPos)
 				posPtr := (*[]int)(*(*unsafe.Pointer)(unsafe.Pointer(&ptr)))
 				*posPtr = append(*posPtr, len(b))
+
 				if idx < length {
 					ptr := load(ctxptr, code.mapIter)
 					iter := e.ptrToUnsafePtr(ptr)
@@ -462,14 +459,15 @@ func (e *Encoder) runEscaped(ctx *encodeRuntimeContext, b []byte, code *opcode)
 		case opMapEnd:
 			// this operation only used by sorted map.
 			length := int(load(ctxptr, code.length))
+			ptr := load(ctxptr, code.mapPos)
+			posPtr := e.ptrToUnsafePtr(ptr)
+			pos := *(*[]int)(posPtr)
 			type mapKV struct {
 				key   string
 				value string
 			}
 			kvs := make([]mapKV, 0, length)
-			ptr := load(ctxptr, code.mapPos)
-			posPtr := e.ptrToUnsafePtr(ptr)
-			pos := *(*[]int)(posPtr)
+
 			for i := 0; i < length; i++ {
 				startKey := pos[i*2]
 				startValue := pos[i*2+1]

encode_vm_escaped_indent.go

@@ -11,11 +11,12 @@ import (
 	"unsafe"
 )
 
-func (e *Encoder) runEscapedIndent(ctx *encodeRuntimeContext, b []byte, code *opcode) ([]byte, error) {
+func (e *Encoder) runEscapedIndent(ctx *encodeRuntimeContext, b []byte, codeSet *opcodeSet) ([]byte, error) {
 	recursiveLevel := 0
 	var seenPtr map[uintptr]struct{}
 	ptrOffset := uintptr(0)
 	ctxptr := ctx.ptr()
+	code := codeSet.code
 
 	for {
 		switch code.op {

encode_vm_indent.go

@@ -11,11 +11,12 @@ import (
 	"unsafe"
 )
 
-func (e *Encoder) runIndent(ctx *encodeRuntimeContext, b []byte, code *opcode) ([]byte, error) {
+func (e *Encoder) runIndent(ctx *encodeRuntimeContext, b []byte, codeSet *opcodeSet) ([]byte, error) {
 	recursiveLevel := 0
 	var seenPtr map[uintptr]struct{}
 	ptrOffset := uintptr(0)
 	ctxptr := ctx.ptr()
+	code := codeSet.code
 
 	for {
 		switch code.op {