avoid some allocs to reduce GC overhead

Author: markus-wa

pkg/demoinfocs/parser.go

@@ -104,6 +104,7 @@ type parser struct {
 	equipmentTypePerModel map[uint64]common.EquipmentType                 // Used to retrieve the EquipmentType of grenade projectiles based on models value. Source 2 only.
 	stringTables          []createStringTable                             // Contains all created sendtables, needed when updating them
 	delayedEventHandlers  []func()                                        // Contains event handlers that need to be executed at the end of a tick (e.g. flash events because FlashDuration isn't updated before that)
+	pendingMessagesCache  []pendingMessage                                // Cache for pending messages that need to be dispatched after the current tick
 }
 
 // NetMessageCreator creates additional net-messages to be dispatched to net-message handlers.

pkg/demoinfocs/s2_commands.go

@@ -310,21 +310,21 @@ func (p *parser) handleDemoPacket(pack *msgs2.CDemoPacket) {
 
 	r := bitread.NewSmallBitReader(bytes.NewReader(b))
 
-	ms := make([]pendingMessage, 0)
+	p.pendingMessagesCache = p.pendingMessagesCache[:0]
 
 	for len(b)*8-r.ActualPosition() > 7 {
 		t := int32(r.ReadUBitInt())
 		size := r.ReadVarInt32()
 		buf := r.ReadBytes(int(size))
 
-		ms = append(ms, pendingMessage{t, buf})
+		p.pendingMessagesCache = append(p.pendingMessagesCache, pendingMessage{t, buf})
 	}
 
-	sort.SliceStable(ms, func(i, j int) bool {
-		return ms[i].priority() < ms[j].priority() // TODO: taken from dotabuff/manta. do we really need this?
+	sort.SliceStable(p.pendingMessagesCache, func(i, j int) bool {
+		return p.pendingMessagesCache[i].priority() < p.pendingMessagesCache[j].priority()
 	})
 
-	for _, m := range ms {
+	for _, m := range p.pendingMessagesCache {
 		var msgCreator NetMessageCreator
 
 		if m.t < int32(msgs2.SVC_Messages_svc_ServerInfo) {

pkg/demoinfocs/sendtables2/entity.go

@@ -484,12 +484,7 @@ func (p *Parser) OnPacketEntities(m *msgs2.CSVCMsg_PacketEntities) error {
 		p.entityFullPackets++
 	}
 
-	type tuple struct {
-		ent *Entity
-		op  st.EntityOp
-	}
-
-	var tuples []tuple
+	p.tuplesCache = p.tuplesCache[:0]
 
 	for ; updates > 0; updates-- {
 		var (
@@ -575,10 +570,10 @@ func (p *Parser) OnPacketEntities(m *msgs2.CSVCMsg_PacketEntities) error {
 			}
 		}
 
-		tuples = append(tuples, tuple{e, op})
+		p.tuplesCache = append(p.tuplesCache, tuple{e, op})
 	}
 
-	for _, t := range tuples {
+	for _, t := range p.tuplesCache {
 		e := t.ent
 
 		for _, h := range p.entityHandlers {

pkg/demoinfocs/sendtables2/field_decoder.go

@@ -260,6 +260,7 @@ func quantizedFactory(f *field) fieldDecoder {
 	}
 
 	qfd := newQuantizedFloatDecoder(f.bitCount, f.encodeFlags, f.lowValue, f.highValue)
+
 	return func(r *reader) interface{} {
 		return qfd.decode(r)
 	}
@@ -274,9 +275,11 @@ func vectorFactory(n int) fieldFactory {
 		d := floatFactory(f)
 		return func(r *reader) interface{} {
 			x := make([]float32, n)
+
 			for i := 0; i < n; i++ {
 				x[i] = d(r).(float32)
 			}
+
 			return x
 		}
 	}
@@ -319,7 +322,7 @@ func ammoDecoder(r *reader) interface{} {
 }
 
 func noscaleDecoder(r *reader) interface{} {
-	return math.Float32frombits(r.readBits(32))
+	return math.Float32frombits(r.readLeUint32())
 }
 
 func runeTimeDecoder(r *reader) interface{} {

pkg/demoinfocs/sendtables2/parser.go

@@ -51,6 +51,11 @@ var itemCounts = map[string]int{
 	"MAX_ABILITY_DRAFT_ABILITIES": 48,
 }
 
+type tuple struct {
+	ent *Entity
+	op  st.EntityOp
+}
+
 type Parser struct {
 	serializers       map[string]*serializer
 	classIdSize       uint32
@@ -61,6 +66,7 @@ type Parser struct {
 	entities          map[int32]*Entity
 	entityHandlers    []st.EntityHandler
 	pathCache         []*fieldPath
+	tuplesCache       []tuple
 }
 
 func (p *Parser) ReadEnterPVS(r *bit.BitReader, index int, entities map[int]st.Entity, slot int) st.Entity {

pkg/demoinfocs/sendtables2/reader.go

@@ -39,13 +39,15 @@ func (r *reader) remBytes() uint32 {
 
 // nextByte reads the next byte from the buffer
 func (r *reader) nextByte() byte {
-	r.pos++
-
-	if r.pos > r.size {
+	if r.pos >= r.size {
 		_panicf("nextByte: insufficient buffer (%d of %d)", r.pos, r.size)
 	}
 
-	return r.buf[r.pos-1]
+	x := r.buf[r.pos]
+
+	r.pos++
+
+	return x
 }
 
 // readBits returns the uint32 value for the given number of sequential bits
@@ -77,22 +79,31 @@ func (r *reader) readBytes(n uint32) []byte {
 	// Fast path if we're byte aligned
 	if r.bitCount == 0 {
 		r.pos += n
+
 		if r.pos > r.size {
 			_panicf("readBytes: insufficient buffer (%d of %d)", r.pos, r.size)
 		}
+
 		return r.buf[r.pos-n : r.pos]
 	}
 
 	buf := make([]byte, n)
+
 	for i := uint32(0); i < n; i++ {
 		buf[i] = byte(r.readBits(8))
 	}
+
 	return buf
 }
 
 // readLeUint32 reads an little-endian uint32
 func (r *reader) readLeUint32() uint32 {
-	return binary.LittleEndian.Uint32(r.readBytes(4))
+	// Fast path if we're byte aligned
+	if r.bitCount == 0 {
+		return binary.LittleEndian.Uint32(r.readBytes(4))
+	}
+
+	return r.readBits(32)
 }
 
 // readLeUint64 reads a little-endian uint64