Merge pull request #213 from jxskiss/compact_type_cache

Add type addrShift to enable bigger encoder/decoder cache

codec.go

@@ -16,6 +16,7 @@ var (
 	cachedDecoderMap unsafe.Pointer // map[uintptr]decoder
 	baseTypeAddr     uintptr
 	maxTypeAddr      uintptr
+	typeAddrShift    uintptr
 )
 
 //go:linkname typelinks reflect.typelinks
@@ -37,6 +38,8 @@ func setupCodec() error {
 	var (
 		min         uintptr = uintptr(^uint(0))
 		max         uintptr = 0
+		isAligned64         = true
+		isAligned32         = true
 	)
 	for i := 0; i < len(offset); i++ {
 		typ := (*rtype)(rtypeOff(section, offset[i]))
@@ -56,15 +59,25 @@ func setupCodec() error {
 				max = addr
 			}
 		}
+
+		// check every address is aligned from the base address
+		isAligned64 = isAligned64 && (addr-min)&63 == 0
+		isAligned32 = isAligned32 && (addr-min)&31 == 0
 	}
 	addrRange := max - min
 	if addrRange == 0 {
 		return fmt.Errorf("failed to get address range of types")
 	}
-	if addrRange > maxAcceptableTypeAddrRange {
+	if isAligned64 {
+		typeAddrShift = 6
+	} else if isAligned32 {
+		typeAddrShift = 5
+	}
+	cacheSize := addrRange >> typeAddrShift
+	if cacheSize > maxAcceptableTypeAddrRange {
 		return fmt.Errorf("too big address range %d", addrRange)
 	}
-	cachedDecoder = make([]decoder, addrRange)
+	cachedDecoder = make([]decoder, cacheSize)
 	baseTypeAddr = min
 	maxTypeAddr = max
 	return nil

codec_go16_test.go

@@ -0,0 +1,13 @@
+// +build go1.16
+
+package json
+
+import (
+	"testing"
+)
+
+func TestTypeAddressAligned32(t *testing.T) {
+	if typeAddrShift != 5 {
+		t.Fatalf("unexpected type address shift %d, want 5", typeAddrShift)
+	}
+}

decode_compile_norace.go

@@ -10,7 +10,7 @@ func decodeCompileToGetDecoder(typ *rtype) (decoder, error) {
 		return decodeCompileToGetDecoderSlowPath(typeptr, typ)
 	}
 
-	index := typeptr - baseTypeAddr
+	index := (typeptr - baseTypeAddr) >> typeAddrShift
 	if dec := cachedDecoder[index]; dec != nil {
 		return dec, nil
 	}

decode_compile_race.go

@@ -15,7 +15,7 @@ func decodeCompileToGetDecoder(typ *rtype) (decoder, error) {
 		return decodeCompileToGetDecoderSlowPath(typeptr, typ)
 	}
 
-	index := typeptr - baseTypeAddr
+	index := (typeptr - baseTypeAddr) >> typeAddrShift
 	decMu.RLock()
 	if dec := cachedDecoder[index]; dec != nil {
 		decMu.RUnlock()

internal/encoder/compiler.go

@@ -27,7 +27,7 @@ func init() {
 	if typeAddr == nil {
 		typeAddr = &runtime.TypeAddr{}
 	}
-	cachedOpcodeSets = make([]*OpcodeSet, typeAddr.AddrRange)
+	cachedOpcodeSets = make([]*OpcodeSet, typeAddr.AddrRange>>typeAddr.AddrShift)
 }
 
 func loadOpcodeMap() map[uintptr]*OpcodeSet {

internal/encoder/compiler_norace.go

@@ -12,7 +12,7 @@ func CompileToGetCodeSet(typeptr uintptr) (*OpcodeSet, error) {
 	if typeptr > typeAddr.MaxTypeAddr {
 		return compileToGetCodeSetSlowPath(typeptr)
 	}
-	index := typeptr - typeAddr.BaseTypeAddr
+	index := (typeptr - typeAddr.BaseTypeAddr) >> typeAddr.AddrShift
 	if codeSet := cachedOpcodeSets[index]; codeSet != nil {
 		return codeSet, nil
 	}

internal/encoder/compiler_race.go

@@ -15,7 +15,7 @@ func CompileToGetCodeSet(typeptr uintptr) (*OpcodeSet, error) {
 	if typeptr > typeAddr.MaxTypeAddr {
 		return compileToGetCodeSetSlowPath(typeptr)
 	}
-	index := typeptr - typeAddr.BaseTypeAddr
+	index := (typeptr - typeAddr.BaseTypeAddr) >> typeAddr.AddrShift
 	setsMu.RLock()
 	if codeSet := cachedOpcodeSets[index]; codeSet != nil {
 		setsMu.RUnlock()

internal/runtime/type.go

@@ -19,6 +19,7 @@ type TypeAddr struct {
 	BaseTypeAddr uintptr
 	MaxTypeAddr  uintptr
 	AddrRange    uintptr
+	AddrShift    uintptr
 }
 
 var (
@@ -51,6 +52,8 @@ func AnalyzeTypeAddr() *TypeAddr {
 	var (
 		min         uintptr = uintptr(^uint(0))
 		max         uintptr = 0
+		isAligned64         = true
+		isAligned32         = true
 	)
 	for i := 0; i < len(offset); i++ {
 		typ := (*Type)(rtypeOff(section, offset[i]))
@@ -70,18 +73,28 @@ func AnalyzeTypeAddr() *TypeAddr {
 				max = addr
 			}
 		}
+		isAligned64 = isAligned64 && (addr-min)&63 == 0
+		isAligned32 = isAligned32 && (addr-min)&31 == 0
 	}
 	addrRange := max - min
 	if addrRange == 0 {
 		return nil
 	}
-	if addrRange > maxAcceptableTypeAddrRange {
+	var addrShift uintptr
+	if isAligned64 {
+		addrShift = 6
+	} else if isAligned32 {
+		addrShift = 5
+	}
+	cacheSize := addrRange >> addrShift
+	if cacheSize > maxAcceptableTypeAddrRange {
 		return nil
 	}
 	typeAddr = &TypeAddr{
 		BaseTypeAddr: min,
 		MaxTypeAddr:  max,
 		AddrRange:    addrRange,
+		AddrShift:    addrShift,
 	}
 	return typeAddr
 }