Some performance improvements for PyPy - #103

Ah, I see some tests failed. I will look intro that. For the moment, ignore the pull-request.

Now it is ready to merge. The speed improvement is about 25% for my testcase. At the moment, pypy is three times slower than the C implementation.

Thanks.
I'm interested in performance difference between bytearray and manual buffering.
But for now, it's looks good to me

I've already looked into using memoryviews on strs instead of strs to prevent copying in _fb_read. That did not improve performance. I will look into byte arrays this week.

asyncio chose bytearray for receive buffer.
But Python 2's bytearray is not optimized for such usecase.

Python 2.7:

In [2]: x = bytearray(10000)

In [3]: %timeit del x[0]; x.append(1)
1000000 loops, best of 3: 327 ns per loop

In [4]: x = bytearray(100000)

In [5]: %timeit del x[0]; x.append(1)
100000 loops, best of 3: 4.78 µs per loop

Python 3.4:

In [1]: x = bytearray(10000)

In [2]: %timeit del x[0]; x.append(1)
10000000 loops, best of 3: 169 ns per loop

In [3]: x = bytearray(100000)

In [4]: %timeit del x[0]; x.append(1)
10000000 loops, best of 3: 167 ns per loop

PyPy 2.3.1:

In [1]: x = bytearray(10000)

In [2]: %timeit del x[0]; x.append(1)
100000 loops, best of 3: 7.28 µs per loop

In [3]: x = bytearray(100000)

In [4]: %timeit del x[0]; x.append(1)
10000 loops, best of 3: 71.5 µs per loop

I reimplemented the buffer using bytearray. It is much slower than the current implementation.

https://github.com/bwesterb/msgpack-python/tree/bytearray