Use Int32x4List to optimize mask/unmask websockets.

sdk/lib/io/websocket_impl.dart

Index: sdk/lib/io/websocket_impl.dart
diff --git a/sdk/lib/io/websocket_impl.dart b/sdk/lib/io/websocket_impl.dart
index 69a7f0eb69f36793cff1db27e937694c2ec5d299..4fc5b72631b57af9460728a56b3f80355e2fc1ba 100644
--- a/sdk/lib/io/websocket_impl.dart
+++ b/sdk/lib/io/websocket_impl.dart
@@ -169,17 +169,15 @@ class _WebSocketProtocolTransformer implements StreamTransformer, EventSink {
         } else {
           assert(_state == PAYLOAD);
           // The payload is not handled one byte at a time but in blocks.
-          int payload = min(lastIndex - index, _remainingPayloadBytes);
-          _remainingPayloadBytes -= payload;
+          int payloadLength = min(lastIndex - index, _remainingPayloadBytes);
+          _remainingPayloadBytes -= payloadLength;
           // Unmask payload if masked.
           if (_masked) {
-            for (int i = index; i < index + payload; i++) {
-              buffer[i] ^= _maskingBytes[_unmaskingIndex++ & 3];
-            }
+            _unmask(index, payloadLength, buffer);
           }
           // Control frame and data frame share _payload builder.
-          _payload.add(new Uint8List.view(buffer.buffer, index, payload));
-          index += payload;
+          _payload.add(new Uint8List.view(buffer.buffer, index, payloadLength));
+          index += payloadLength;
           if (_isControlFrame()) {
             if (_remainingPayloadBytes == 0) _controlFrameEnd();
           } else {
@@ -200,6 +198,40 @@ class _WebSocketProtocolTransformer implements StreamTransformer, EventSink {
     }
   }
 
+  void _unmask(int index, int payload, Uint8List buffer) {

[kasperl, 2014/01/29 09:47:37]

payload -> length

[Anders Johnsen, 2014/01/29 10:19:04]

Done.

+    const int BLOCK_SIZE = 16;
+    // Start by alligning to 16 bytes.

[kasperl, 2014/01/29 09:47:37]

alligning -> aligning

[Anders Johnsen, 2014/01/29 10:19:04]

Done.

+    final int startOffset = min(BLOCK_SIZE - (index & 15), payload);

[kasperl, 2014/01/29 09:47:37]

Maybe replace this whole min computation by checking if the length is less than
BLOCK_SIZE. If it is, you can skip the alignment part and the block handling
part completely.

[Anders Johnsen, 2014/01/29 10:19:04]

Done.

+    int end = index + startOffset;
+    for (int i = index; i < end; i++) {
+      buffer[i] ^= _maskingBytes[_unmaskingIndex++ & 3];
+    }
+    index += startOffset;
+    payload -= startOffset;
+    final int blockCount = payload ~/ BLOCK_SIZE;
+    if (blockCount > 0) {
+      // Create mask block.
+      int mask = 0;
+      for (int i = 3; i >= 0; i--) {

[kasperl, 2014/01/29 09:47:37]

Here's an idea: Why don't you turn masking bytes into a Uint8List and make it
contain the following bytes:

   [ m0, m1, m2, m3,
     m0, m1, m2, m3,
     m0, m1, m2, m3,
     m0, m1, m2, m3,
     
     m1, m2, m3, m0,
     m1, m2, m3, m0,
     m1, m2, m3, m0,
     m1, m2, m3, m0,

     m2, m3, m0, m1,
     m2, m3, m0, m1,
     m2, m3, m0, m1,
     m2, m3, m0, m1,

     m3, m0, m1, m2,
     m3, m0, m1, m2,
     m3, m0, m1, m2,
     m3, m0, m1, m2 ]

You can fill this information in when you read the masking bytes before calling
maskDone(). If you create a Int32x4List view of these bytes and fetch the block
mask from it using:

   Int32x4 blockMask = _maskingBlocks[_unmaskingIndex];

Feel free to experiment with this in another CL.

[Anders Johnsen, 2014/01/29 10:19:04]

While the idea is very cool, I'm afraid the build-overhead may be too expensive
in the case of small packages. I'll keep as is, and investigate further in
another Cl.

+        mask = (mask << 8) | _maskingBytes[(_unmaskingIndex + i) & 3];
+      }
+      Int32x4 blockMask = new Int32x4(mask, mask, mask, mask);
+      Int32x4List blockBuffer = new Int32x4List.view(
+          buffer.buffer, index, blockCount);
+      for (int i = 0; i < blockBuffer.length; i++) {
+        blockBuffer[i] ^= blockMask;
+      }
+      final int bytes = blockCount * BLOCK_SIZE;
+      index += bytes;
+      payload -= bytes;
+    }
+    // Handle end.
+    end = index + payload;
+    for (int i = index; i < end; i++) {
+      buffer[i] ^= _maskingBytes[_unmaskingIndex++ & 3];
+    }
+  }
+
   void _lengthDone() {
     if (_masked) {
       if (!_serverSide) {
@@ -530,28 +562,45 @@ class _WebSocketOutgoingTransformer implements StreamTransformer, EventSink {
       header.setRange(index, index + 4, maskBytes);
       index += 4;
       if (data != null) {
-        var list;
+        Uint8List list;
         // If this is a text message just do the masking inside the
         // encoded data.
-        if (opcode == _WebSocketOpcode.TEXT) {
+        if (opcode == _WebSocketOpcode.TEXT && data is Uint8List) {
           list = data;
         } else {
-          list = new Uint8List(data.length);
+          if (data is Uint8List) {
+            list = new Uint8List.fromList(data);
+          } else {
+            list = new Uint8List(data.length);
+            for (int i = 0; i < data.length; i++) {
+              if (data[i] < 0 || 255 < data[i]) {
+                throw new ArgumentError(
+                    "List element is not a byte value "
+                    "(value ${data[i]} at index $i)");
+              }
+              list[i] = data[i];
+            }
+          }
         }
-        if (data is Uint8List) {
-          for (int i = 0; i < data.length; i++) {
-            list[i] = data[i] ^ maskBytes[i & 3];
+        const int BLOCK_SIZE = 16;
+        int blockCount = list.length ~/ BLOCK_SIZE;
+        if (blockCount > 0) {
+          // Create mask block.
+          int mask = 0;
+          for (int i = 3; i >= 0; i--) {
+            mask = (mask << 8) | maskBytes[i];
           }
-        } else {
-          for (int i = 0; i < data.length; i++) {
-            if (data[i] < 0 || 255 < data[i]) {
-              throw new ArgumentError(
-                  "List element is not a byte value "
-                  "(value ${data[i]} at index $i)");
-            }
-            list[i] = data[i] ^ maskBytes[i & 3];
+          Int32x4 blockMask = new Int32x4(mask, mask, mask, mask);
+          Int32x4List blockBuffer = new Int32x4List.view(
+              list.buffer, 0, blockCount);
+          for (int i = 0; i < blockBuffer.length; i++) {
+            blockBuffer[i] ^= blockMask;
           }
         }
+        // Handle end.
+        for (int i = blockCount * BLOCK_SIZE; i < list.length; i++) {
+          list[i] ^= maskBytes[i & 3];
+        }
         data = list;
       }
     }