为什么反射Type.Implements（）比类型断言这么慢？

Question

I'm trying to efficiently test whether an interface{} implements a given function and my solution is to create an interface with just this function and then check whether the interface{} implements this single function interface. The two options here seem to be either using reflection or a type assertion. Both seem to have identical behaviour however there is a large speed difference.

Looking at the code for Value.Implements() it does a linear scan over the functions defined on the value and compares them against the interface. The type assertion however just seems to do a constant time comparison (independent of the number of functions in the interface).

Is there a reason why Implements() doesn't just do a type assertion?

Benchmark:

package benchmarks

import (
    "reflect"
    "testing"
)

type ITest interface {
    Foo()
}

type Base struct{}

func (Base) A() {}
func (Base) B() {}
func (Base) C() {}
func (Base) D() {}
func (Base) E() {}
func (Base) F() {}
func (Base) G() {}
func (Base) H() {}
func (Base) I() {}
func (Base) J() {}

var Interface = reflect.TypeOf((*ITest)(nil)).Elem()

func BenchmarkReflection(b *testing.B) {
    var iface interface{}
    iface = Base{}
    for i := 0; i < b.N; i++ {
        if reflect.TypeOf(iface).Implements(Interface) {
            b.FailNow()
        }
    }
}

func BenchmarkAssertion(b *testing.B) {
    var iface interface{}
    iface = Base{}
    for i := 0; i < b.N; i++ {
        if _, ok := iface.(ITest); ok {
            b.FailNow()
        }
    }
}

Results:

go test -run=XXX -bench=. so_test.go
goos: linux
goarch: amd64
BenchmarkReflection-8           10000000                  208 ns/op
BenchmarkAssertion-8            200000000                9.24 ns/op
PASS
ok      command-line-arguments  5.115s

Answer 1

Type assertion in Go relies on a function called runtime.assertE2I2. If you look into the code, you'll notice it relies on getitab which in turn relies on additab (in the same file).

Now, the actual logic of checking if the given type implements an interface inside of additab is exactly the same as Implements in reflect - a linear search, which is even pointed out in this comment:

// both inter and typ have method sorted by name,
// and interface names are unique,
// so can iterate over both in lock step;
// the loop is O(ni+nt) not O(ni*nt).

However, the difference is that additab actually utilises caching - the result of the type assertion is stored in a hash map, so subsequent type assertions for the same type will run in constant time, which is why you're seeing a huge difference in performance.