http2: allocate buffer pools using pointers to arrays

This remove the allocation for the slice header in sync.Pool.New and putDataBufferChunk.
It also divide the number of allocations kept alive by the pool.

Change-Id: Icf493ebc568ae80a4e73e9768a6f1c7fce8e1365
Reviewed-on: https://go-review.googlesource.com/c/net/+/539915
Reviewed-by: Bryan Mills <[email protected]>
Auto-Submit: Damien Neil <[email protected]>
LUCI-TryBot-Result: Go LUCI <[email protected]>
Reviewed-by: Brad Fitzpatrick <[email protected]>
Reviewed-by: Damien Neil <[email protected]>
TryBot-Result: Gopher Robot <[email protected]>
Run-TryBot: qiulaidongfeng <[email protected]>
Reviewed-by: qiulaidongfeng <[email protected]>

http2/databuffer.go

@@ -20,41 +20,44 @@ import (
 // TODO: Benchmark to determine if the pools are necessary. The GC may have
 // improved enough that we can instead allocate chunks like this:
 // make([]byte, max(16<<10, expectedBytesRemaining))
-var (
-	dataChunkSizeClasses = []int{
-		1 << 10,
-		2 << 10,
-		4 << 10,
-		8 << 10,
-		16 << 10,
-	}
-	dataChunkPools = [...]sync.Pool{
-		{New: func() interface{} { return make([]byte, 1<<10) }},
-		{New: func() interface{} { return make([]byte, 2<<10) }},
-		{New: func() interface{} { return make([]byte, 4<<10) }},
-		{New: func() interface{} { return make([]byte, 8<<10) }},
-		{New: func() interface{} { return make([]byte, 16<<10) }},
-	}
-)
+var dataChunkPools = [...]sync.Pool{
+	{New: func() interface{} { return new([1 << 10]byte) }},
+	{New: func() interface{} { return new([2 << 10]byte) }},
+	{New: func() interface{} { return new([4 << 10]byte) }},
+	{New: func() interface{} { return new([8 << 10]byte) }},
+	{New: func() interface{} { return new([16 << 10]byte) }},
+}
 
 func getDataBufferChunk(size int64) []byte {
-	i := 0
-	for ; i < len(dataChunkSizeClasses)-1; i++ {
-		if size <= int64(dataChunkSizeClasses[i]) {
-			break
-		}
+	switch {
+	case size <= 1<<10:
+		return dataChunkPools[0].Get().(*[1 << 10]byte)[:]
+	case size <= 2<<10:
+		return dataChunkPools[1].Get().(*[2 << 10]byte)[:]
+	case size <= 4<<10:
+		return dataChunkPools[2].Get().(*[4 << 10]byte)[:]
+	case size <= 8<<10:
+		return dataChunkPools[3].Get().(*[8 << 10]byte)[:]
+	default:
+		return dataChunkPools[4].Get().(*[16 << 10]byte)[:]
 	}
-	return dataChunkPools[i].Get().([]byte)
 }
 
 func putDataBufferChunk(p []byte) {
-	for i, n := range dataChunkSizeClasses {
-		if len(p) == n {
-			dataChunkPools[i].Put(p)
-			return
-		}
+	switch len(p) {
+	case 1 << 10:
+		dataChunkPools[0].Put((*[1 << 10]byte)(p))
+	case 2 << 10:
+		dataChunkPools[1].Put((*[2 << 10]byte)(p))
+	case 4 << 10:
+		dataChunkPools[2].Put((*[4 << 10]byte)(p))
+	case 8 << 10:
+		dataChunkPools[3].Put((*[8 << 10]byte)(p))
+	case 16 << 10:
+		dataChunkPools[4].Put((*[16 << 10]byte)(p))
+	default:
+		panic(fmt.Sprintf("unexpected buffer len=%v", len(p)))
 	}
-	panic(fmt.Sprintf("unexpected buffer len=%v", len(p)))
 }
 
 // dataBuffer is an io.ReadWriter backed by a list of data chunks.