From 8448545f14ff3e9be0b1cbea6a7804a454ef0cea Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=D0=9A=D0=BE=D1=80=D0=B5=D0=BD=D0=B1=D0=B5=D1=80=D0=B3=20?=
 =?UTF-8?q?=D0=9C=D0=B0=D1=80=D0=BA=20=28imac=29?= <socketpair@gmail.com>
Date: Thu, 17 Dec 2015 15:51:16 +0500
Subject: [PATCH] Weboscket XOR performance improved. Fixes #686

Both cython and pure python-based.
---
 aiohttp/_websocket.pyx | 52 ++++++++++++++++++++++++++++++++----------
 aiohttp/websocket.py   | 14 +++++++++++-
 2 files changed, 53 insertions(+), 13 deletions(-)

diff --git a/aiohttp/_websocket.pyx b/aiohttp/_websocket.pyx
index 8e74ac134d2..a2d2c55f20c 100644
--- a/aiohttp/_websocket.pyx
+++ b/aiohttp/_websocket.pyx
@@ -1,20 +1,48 @@
-from cpython cimport PyBytes_FromStringAndSize, PyBytes_AsString
-from cpython.ref cimport PyObject
+from cpython cimport PyBytes_AsString
 
+#from cpython cimport PyByteArray_AsString # cython still not exports that
 cdef extern from "Python.h":
-    char* PyByteArray_AsString(object bytearray) except NULL
+    char* PyByteArray_AsString(bytearray ba) except NULL
 
+from libc.stdint cimport uint32_t, uint64_t, uintmax_t
 
 def _websocket_mask_cython(bytes mask, bytearray data):
-    cdef Py_ssize_t mask_len, data_len, i
-    cdef char * in_buf
-    cdef char * out_buf
-    cdef char * mask_buf
-    cdef bytes ret
-    mask_len = len(mask)
+    """Note, this function mutates it's `data` argument
+    """
+    cdef:
+        Py_ssize_t data_len, i
+        # bit operations on signed integers are implementation-specific
+        unsigned char * in_buf
+        const unsigned char * mask_buf
+        uint32_t uint32_msk
+        uint64_t uint64_msk
+
+    assert len(mask) == 4
+
     data_len = len(data)
-    in_buf = PyByteArray_AsString(data)
-    mask_buf = PyBytes_AsString(mask)
+    in_buf = <unsigned char*>PyByteArray_AsString(data)
+    mask_buf = <const unsigned char*>PyBytes_AsString(mask)
+    uint32_msk = (<uint32_t*>mask_buf)[0]
+
+    # TODO: align in_data ptr to achieve even faster speeds
+    # does it need in python ?! malloc() always aligns to sizeof(long) bytes
+
+    if sizeof(uintmax_t) >= 8:
+        uint64_msk = uint32_msk
+        uint64_msk = (uint64_msk << 32) | uint32_msk
+
+        while data_len >= 8:
+            (<uint64_t*>in_buf)[0] ^= uint64_msk
+            in_buf += 8
+            data_len -= 8
+
+
+    while data_len >= 4:
+        (<uint32_t*>in_buf)[0] ^= uint32_msk
+        in_buf += 4
+        data_len -= 4
+
     for i in range(0, data_len):
-        in_buf[i] = in_buf[i] ^ mask_buf[i % 4]
+        in_buf[i] ^= mask_buf[i]
+
     return data
diff --git a/aiohttp/websocket.py b/aiohttp/websocket.py
index 10056464e31..1853b6eac2c 100644
--- a/aiohttp/websocket.py
+++ b/aiohttp/websocket.py
@@ -6,6 +6,7 @@
 import hashlib
 import os
 import random
+import sys
 
 from struct import Struct
 from aiohttp import errors, hdrs
@@ -172,6 +173,9 @@ def WebSocketParser(out, buf):
                     Message(OPCODE_BINARY, data, ''), len(data))
 
 
+native_byteorder = sys.byteorder
+
+
 def _websocket_mask_python(mask, data):
     """Websocket masking function.
 
@@ -184,7 +188,15 @@ def _websocket_mask_python(mask, data):
     version when available.
 
     """
-    return bytes(b ^ mask[i % 4] for i, b in enumerate(data))
+    assert len(mask) == 4
+    datalen = len(data)
+    if datalen == 0:
+        # everything work without this, but may be changed later in Python.
+        return b''
+    data = int.from_bytes(data, native_byteorder)
+    mask = int.from_bytes(mask * (datalen // 4) + mask[: datalen % 4],
+                          native_byteorder)
+    return (data ^ mask).to_bytes(datalen, native_byteorder)
 
 
 if bool(os.environ.get('AIOHTTP_NO_EXTENSIONS')):