GO 反射对性能的影响分析

本文介绍了 GO 反射对性能的影响分析

今天在公司写了一段代码，判断一个变量是否为空值，由于判断的类型太少，code review的时候同事说还有很多类型没有考虑到，并且提到有没有开源的包做这个事，于是找了一段assert.IsEmpty里面的代码。但是这段代码用到了反射。在code review的时候同事又提到了反射影响性能。

基于此，这里对比了一下两种方式实习IsEmpty的性能问题。废话不多说，上代码。

代码

最开始的代码

func IsEmpty(val interface{}) bool {
	if val == nil {
		return true
	}
	switch v := val.(type) {
	case int:
		return v == int(0)
	case int8:
		return v == int8(0)
	case int16:
		return v == int16(0)
	case int32:
		return v == int32(0)
	case int64:
		return v == int64(0)
	case string:
		return v == ""
	default:
		return false
	}
}

由于目前场景里面只需要判断这几种数据类型，因此只实现了几种。这种做法明细很不好，将来如果有别人用怎么办？这是一种偷懒的做法，不是一个高级程序员应该有的素质。

在同事提出可能会有其他数据类型的时候，想着类型太多，穷举容易漏，于是在网上找了一下开源的包，遗憾没有找到。但是想到了assert.Empty函数，于是看了一下源代码，找到了它的实现方法。

// isEmpty gets whether the specified object is considered empty or not.
func isEmpty(object interface{}) bool {
	// get nil case out of the way
	if object == nil {
		return true
	}
	objValue := reflect.ValueOf(object)
	switch objValue.Kind() {
	// collection types are empty when they have no element
	case reflect.Array, reflect.Chan, reflect.Map, reflect.Slice:
		return objValue.Len() == 0
		// pointers are empty if nil or if the value they point to is empty
	case reflect.Ptr:
		if objValue.IsNil() {
			return true
		}
		deref := objValue.Elem().Interface()
		return isEmpty(deref)
		// for all other types, compare against the zero value
	default:
		zero := reflect.Zero(objValue.Type())
		return reflect.DeepEqual(object, zero.Interface())
	}
}

于是把这段代码复制过来，code review的时候同事说反射会影响性能（其实我觉得还好，不知道大家觉得它对性能影响有多大，欢迎留言讨论），于是我又改了一版。结合了上面两种方法。先判断是不是基础数据类型，如果不是再用反射。

const (
	IntZero     = int(0)
	Int8Zero    = int8(0)
	Int16Zero   = int16(0)
	Int32Zero   = int32(0)
	Int64Zero   = int64(0)
	UintZero    = uint(0)
	Uint8Zero   = uint8(0)
	Uint16Zero  = uint16(0)
	Uint32Zero  = uint32(0)
	Uint64Zero  = uint64(0)
	Float32Zero = float32(0)
	Float64Zero = float64(0)
	StringZero  = ""
)
func IsEmpty(data interface{}) bool {
	if data == nil {
		return false
	}
	switch v := data.(type) {
	case bool:
		return false
	case *bool:
		return v == nil
	case int:
		return v == IntZero
	case *int:
		return v == nil || *v == IntZero
	case int8:
		return v == Int8Zero
	case *int8:
		return v == nil || *v == Int8Zero
	case int16:
		return v == Int16Zero
	case *int16:
		return v == nil || *v == Int16Zero
	case int32:
		return v == Int32Zero
	case *int32:
		return v == nil || *v == Int32Zero
	case int64:
		return v == Int64Zero
	case *int64:
		return v == nil || *v == Int64Zero
	case uint:
		return v == UintZero
	case *uint:
		return v == nil || *v == UintZero
	case uint8:
		return v == Uint8Zero
	case *uint8:
		return v == nil || *v == Uint8Zero
	case uint16:
		return v == Uint16Zero
	case *uint16:
		return v == nil || *v == Uint16Zero
	case uint32:
		return v == Uint32Zero
	case *uint32:
		return v == nil || *v == Uint32Zero
	case uint64:
		return v == Uint64Zero
	case *uint64:
		return v == nil || *v == Uint64Zero
	case float32:
		return v == Float32Zero
	case *float32:
		return v == nil || *v == Float32Zero
	case float64:
		return v == Float64Zero
	case *float64:
		return v == nil || *v == Float64Zero
	case string:
		return v == StringZero
	case *string:
		return v == nil || *v == StringZero
	default:
		kind := reflect.TypeOf(data).Kind()
		if kind == reflect.Ptr {
			dataKind := reflect.ValueOf(data).Elem().Kind()
			if dataKind == reflect.Invalid || dataKind == reflect.Struct {
				return reflect.ValueOf(data).IsNil()
			} else {
				return false
			}
		} else if kind == reflect.Slice || kind == reflect.Map {
			// slice
			return reflect.ValueOf(data).Len() == 0
		} else if kind == reflect.Struct {
			// struct
			return false
		} else {
			panic("not support type. you can support by yourself")
		}
	}
}

得到第三版。

性能分析

代码提交之后我一直在想第一版和第二版性能到底差别有多大。其实这个时候我偷懒了，没有做性能分析，做个Bench分析一下很简单，这件事一直在我心里，过了一天终于写了一段代码比对一下性能。

func BenchmarkTestIsEmpty(t *testing.B) {
	v1 := false
	v2 := true
	var v3 *bool
	v4 := int(0)
	var v5 *int
	v6 := int64(0)
	var v7 *int64
	v8 := ""
	var v9 *string
	v10 := "test"
	v11 := float64(0.00)
	v12 := float64(0.01)
	testCases := []TestCase{
		{input: v1, want: false},
		{input: &v1, want: false},
		{input: v2, want: false},
		{input: &v2, want: false},
		{input: v3, want: true},
		{input: v4, want: true},
		{input: &v4, want: true},
		{input: v5, want: true},
		{input: int(1), want: false},
		{input: v6, want: true},
		{input: &v6, want: true},
		{input: v7, want: true},
		{input: int64(1), want: false},
		{input: v8, want: true},
		{input: &v8, want: true},
		{input: v9, want: true},
		{input: v10, want: false},
		{input: &v10, want: false},
		{input: v11, want: true},
		{input: &v11, want: true},
		{input: v12, want: false},
		{input: &v12, want: false},
	}
	for i, testCase := range testCases {
		result := IsEmpty(testCases[i].input)
		assert.Equal(t, testCase.want, result)
	}
}
func BenchmarkTestIsEmptyV1(t *testing.B) {
	v1 := false
	v2 := true
	var v3 *bool
	v4 := int(0)
	var v5 *int
	v6 := int64(0)
	var v7 *int64
	v8 := ""
	var v9 *string
	v10 := "test"
	v11 := float64(0.00)
	v12 := float64(0.01)
	testCases := []TestCase{
		{input: v1, want: true},
		{input: &v1, want: true},
		{input: v2, want: false},
		{input: &v2, want: false},
		{input: v3, want: true},
		{input: v4, want: true},
		{input: &v4, want: true},
		{input: v5, want: true},
		{input: int(1), want: false},
		{input: v6, want: true},
		{input: &v6, want: true},
		{input: v7, want: true},
		{input: int64(1), want: false},
		{input: v8, want: true},
		{input: &v8, want: true},
		{input: v9, want: true},
		{input: v10, want: false},
		{input: &v10, want: false},
		{input: v11, want: true},
		{input: &v11, want: true},
		{input: v12, want: false},
		{input: &v12, want: false},
	}
	for i, testCase := range testCases {
		result := IsEmptyV1(testCases[i].input)
		assert.Equal(t, testCase.want, result)
	}
}

运行

go test -bench=.  -benchmem

结果

goos: darwin
goarch: amd64
pkg: common/util
cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
BenchmarkTestIsEmpty-12 120820 9635 ns/op 0 B/op 0 allocs/op
BenchmarkTestIsEmptyV1-12 103664 11582 ns/op 72 B/op 8 allocs/op
PASS
ok common/util 5.149s

ns:每次运行耗费的世界。时间复杂的相差10倍
B:每次运行分配的字节数。可见非反射版不需要额外的内存
allocs:每次运行分配内存次数。

综上可见，性能差别挺大的。如果只看运行时间，相差了10倍。确实反射影响性能，以后还是少用为好，最好不要在循环里面用。

最终解决办法

最终结合第一版和第二版，写出了第三版。

catroom

代码

性能分析

推荐文章

发表评论

catroom