encoding/binary: cache struct sizes to speed up Read and Write for slice of structs.

A lot of allocations happen in dataSize due to reflection.

Cache the result of the function when encoding a
slice of structs similar to what is done for struct types
so that subsequent calls to dataSize can avoid allocations.

                        │   old.txt   │            new.txt            │
                        │   sec/op    │   sec/op     vs base          │
WriteSlice1000Structs-2   846.7µ ± 4%   856.4µ ± 3%  ~ (p=0.602 n=20)

                        │   old.txt    │            new.txt             │
                        │     B/s      │     B/s       vs base          │
WriteSlice1000Structs-2   84.48Mi ± 4%   83.52Mi ± 3%  ~ (p=0.602 n=20)

                        │   old.txt    │               new.txt               │
                        │     B/op     │     B/op      vs base               │
WriteSlice1000Structs-2   80.18Ki ± 0%   80.06Ki ± 0%  -0.15% (p=0.000 n=20)

                        │   old.txt   │              new.txt               │
                        │  allocs/op  │ allocs/op   vs base                │
WriteSlice1000Structs-2   16.000 ± 0%   1.000 ± 0%  -93.75% (p=0.000 n=2

                       │   old.txt   │              new.txt               │
                       │   sec/op    │   sec/op     vs base               │
ReadSlice1000Structs-2   847.4µ ± 4%   821.1µ ± 3%  -3.10% (p=0.012 n=20)

                       │   old.txt    │               new.txt               │
                       │     B/s      │     B/s       vs base               │
ReadSlice1000Structs-2   84.40Mi ± 4%   87.11Mi ± 3%  +3.20% (p=0.012 n=20)

                       │   old.txt    │               new.txt               │
                       │     B/op     │     B/op      vs base               │
ReadSlice1000Structs-2   80.12Ki ± 0%   80.00Ki ± 0%  -0.15% (p=0.000 n=20)

                       │   old.txt   │              new.txt               │
                       │  allocs/op  │ allocs/op   vs base                │
ReadSlice1000Structs-2   16.000 ± 0%   1.000 ± 0%  -93.75% (p=0.000 n=20)

Fixes #66253

Change-Id: I8227e61306db1fe103489ea4fee2429247c3debc
Reviewed-on: https://go-review.googlesource.com/c/go/+/570855
LUCI-TryBot-Result: Go LUCI <golang-scoped@luci-project-accounts.iam.gserviceaccount.com>
Reviewed-by: Ian Lance Taylor <iant@google.com>
Reviewed-by: Cherry Mui <cherryyz@google.com>
Auto-Submit: Ian Lance Taylor <iant@google.com>

src/encoding/binary/binary.go

@@ -480,8 +480,17 @@ var structSize sync.Map // map[reflect.Type]int
 func dataSize(v reflect.Value) int {
 	switch v.Kind() {
 	case reflect.Slice:
-		if s := sizeof(v.Type().Elem()); s >= 0 {
+		t := v.Type().Elem()
-			return s * v.Len()
+		if size, ok := structSize.Load(t); ok {
+			return size.(int) * v.Len()
+		}
+
+		size := sizeof(t)
+		if size >= 0 {
+			if t.Kind() == reflect.Struct {
+				structSize.Store(t, size)
+			}
+			return size * v.Len()
 		}
 
 	case reflect.Struct:

src/encoding/binary/binary_test.go

@@ -631,6 +631,31 @@ func BenchmarkWriteStruct(b *testing.B) {
 	}
 }
 
+func BenchmarkWriteSlice1000Structs(b *testing.B) {
+	slice := make([]Struct, 1000)
+	buf := new(bytes.Buffer)
+	var w io.Writer = buf
+	b.SetBytes(int64(Size(slice)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		buf.Reset()
+		Write(w, BigEndian, slice)
+	}
+	b.StopTimer()
+}
+
+func BenchmarkReadSlice1000Structs(b *testing.B) {
+	bsr := &byteSliceReader{}
+	slice := make([]Struct, 1000)
+	buf := make([]byte, Size(slice))
+	b.SetBytes(int64(len(buf)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		bsr.remain = buf
+		Read(bsr, BigEndian, slice)
+	}
+}
+
 func BenchmarkReadInts(b *testing.B) {
 	var ls Struct
 	bsr := &byteSliceReader{}