Optimize Patricia tree

Remove all the extra allocations in the Patricia tree.

O(1) allocation for Patricia and  ~10% improvement for HTTP1 matching.

benchmark                      old ns/op     new ns/op     delta
BenchmarkCMuxConnHTTP1-4       908           782           -13.88%
BenchmarkCMuxConnHTTP2-4       835           818           -2.04%
BenchmarkCMuxConnHTTP1n2-4     1074          1033          -3.82%
BenchmarkCMuxConnHTTP2n1-4     1010          901           -10.79%

benchmark                      old allocs     new allocs     delta
BenchmarkCMuxConnHTTP1-4       5              3              -40.00%
BenchmarkCMuxConnHTTP2-4       4              4              +0.00%
BenchmarkCMuxConnHTTP1n2-4     6              4              -33.33%
BenchmarkCMuxConnHTTP2n1-4     6              4              -33.33%

benchmark                      old bytes     new bytes     delta
BenchmarkCMuxConnHTTP1-4       276           272           -1.45%
BenchmarkCMuxConnHTTP2-4       304           304           +0.00%
BenchmarkCMuxConnHTTP1n2-4     306           304           -0.65%
BenchmarkCMuxConnHTTP2n1-4     308           304           -1.30%

bench_test.go

@@ -6,8 +6,20 @@ import (
 	"net"
 	"sync"
 	"testing"
+
+	"golang.org/x/net/http2"
 )
 
+var (
+	benchHTTP1Payload = make([]byte, 4096)
+	benchHTTP2Payload = make([]byte, 4096)
+)
+
+func init() {
+	copy(benchHTTP1Payload, []byte("GET http://www.w3.org/ HTTP/1.1"))
+	copy(benchHTTP2Payload, http2.ClientPreface)
+}
+
 type mockConn struct {
 	net.Conn
 	r io.Reader
@@ -17,20 +29,19 @@ func (c *mockConn) Read(b []byte) (n int, err error) {
 	return c.r.Read(b)
 }
 
-func BenchmarkCMuxConn(b *testing.B) {
-	benchHTTPPayload := make([]byte, 4096)
-	copy(benchHTTPPayload, []byte("GET http://www.w3.org/ HTTP/1.1"))
+func discard(l net.Listener) {
+	for {
+		if _, err := l.Accept(); err != nil {
+			return
+		}
+	}
+}
 
+func BenchmarkCMuxConnHTTP1(b *testing.B) {
 	m := New(nil).(*cMux)
 	l := m.Match(HTTP1Fast())
 
-	go func() {
-		for {
-			if _, err := l.Accept(); err != nil {
-				return
-			}
-		}
-	}()
+	go discard(l)
 
 	donec := make(chan struct{})
 	var wg sync.WaitGroup
@@ -39,7 +50,67 @@ func BenchmarkCMuxConn(b *testing.B) {
 	b.ResetTimer()
 	for i := 0; i < b.N; i++ {
 		c := &mockConn{
-			r: bytes.NewReader(benchHTTPPayload),
+			r: bytes.NewReader(benchHTTP1Payload),
+		}
+		m.serve(c, donec, &wg)
+	}
+}
+
+func BenchmarkCMuxConnHTTP2(b *testing.B) {
+	m := New(nil).(*cMux)
+	l := m.Match(HTTP2())
+	go discard(l)
+
+	donec := make(chan struct{})
+	var wg sync.WaitGroup
+	wg.Add(b.N)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		c := &mockConn{
+			r: bytes.NewReader(benchHTTP2Payload),
+		}
+		m.serve(c, donec, &wg)
+	}
+}
+
+func BenchmarkCMuxConnHTTP1n2(b *testing.B) {
+	m := New(nil).(*cMux)
+	l1 := m.Match(HTTP1Fast())
+	l2 := m.Match(HTTP2())
+
+	go discard(l1)
+	go discard(l2)
+
+	donec := make(chan struct{})
+	var wg sync.WaitGroup
+	wg.Add(b.N)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		c := &mockConn{
+			r: bytes.NewReader(benchHTTP2Payload),
+		}
+		m.serve(c, donec, &wg)
+	}
+}
+
+func BenchmarkCMuxConnHTTP2n1(b *testing.B) {
+	m := New(nil).(*cMux)
+	l2 := m.Match(HTTP2())
+	l1 := m.Match(HTTP1Fast())
+
+	go discard(l1)
+	go discard(l2)
+
+	donec := make(chan struct{})
+	var wg sync.WaitGroup
+	wg.Add(b.N)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		c := &mockConn{
+			r: bytes.NewReader(benchHTTP1Payload),
 		}
 		m.serve(c, donec, &wg)
 	}

patricia.go

@@ -9,11 +9,19 @@ import (
 // and cannot be changed after instantiation.
 type patriciaTree struct {
 	root *ptNode
+	buf  []byte // preallocated buffer to read data while matching
 }
 
-func newPatriciaTree(b ...[]byte) *patriciaTree {
+func newPatriciaTree(bs ...[]byte) *patriciaTree {
+	max := 0
+	for _, b := range bs {
+		if max < len(b) {
+			max = len(b)
+		}
+	}
 	return &patriciaTree{
-		root: newNode(b),
+		root: newNode(bs),
+		buf:  make([]byte, max+1),
 	}
 }
 
@@ -22,17 +30,17 @@ func newPatriciaTreeString(strs ...string) *patriciaTree {
 	for i, s := range strs {
 		b[i] = []byte(s)
 	}
-	return &patriciaTree{
-		root: newNode(b),
-	}
+	return newPatriciaTree(b...)
 }
 
 func (t *patriciaTree) matchPrefix(r io.Reader) bool {
-	return t.root.match(r, true)
+	n, _ := io.ReadFull(r, t.buf)
+	return t.root.match(t.buf[:n], true)
 }
 
 func (t *patriciaTree) match(r io.Reader) bool {
-	return t.root.match(r, false)
+	n, _ := io.ReadFull(r, t.buf)
+	return t.root.match(t.buf[:n], false)
 }
 
 type ptNode struct {
@@ -122,52 +130,30 @@ func splitPrefix(bss [][]byte) (prefix []byte, rest [][]byte) {
 	return prefix, rest
 }
 
-func readBytes(r io.Reader, n int) (b []byte, err error) {
-	b = make([]byte, n)
-	o := 0
-	for o < n {
-		nr, err := r.Read(b[o:])
-		if err != nil && err != io.EOF {
-			return b, err
+func (n *ptNode) match(b []byte, prefix bool) bool {
+	l := len(n.prefix)
+	if l > 0 {
+		if l > len(b) {
+			l = len(b)
 		}
-
-		o += nr
-
-		if err == io.EOF {
-			break
-		}
-	}
-	return b[:o], nil
-}
-
-func (n *ptNode) match(r io.Reader, prefix bool) bool {
-	if l := len(n.prefix); l > 0 {
-		b, err := readBytes(r, l)
-		if err != nil || len(b) != l || !bytes.Equal(b, n.prefix) {
+		if !bytes.Equal(b[:l], n.prefix) {
 			return false
 		}
 	}
 
-	if prefix && n.terminal {
+	if n.terminal && (prefix || len(n.prefix) == len(b)) {
 		return true
 	}
 
-	b := make([]byte, 1)
-	for {
-		nr, err := r.Read(b)
-		if nr != 0 {
-			break
-		}
-
-		if err == io.EOF {
-			return n.terminal
-		}
-
-		if err != nil {
-			return false
-		}
+	nextN, ok := n.next[b[l]]
+	if !ok {
+		return false
 	}
 
-	nextN, ok := n.next[b[0]]
-	return ok && nextN.match(r, prefix)
+	if l == len(b) {
+		b = b[l:l]
+	} else {
+		b = b[l+1:]
+	}
+	return nextN.match(b, prefix)
 }