Improve Python performance by using MemoryView

[aronsky]

Python bindings for analysis and disassembly plugins passed buffers by using Py_BuildValue and passing the buffer as a string for Python consumption. This required Python to copy the whole buffer - a costly operation, especially when the buffer was large enough.

For a vast majority of architectures, copying the whole buffer is unnecessary, as the opcode can't be longer than several bytes (e.g., on x86 an opcode's size is limited to 15 bytes).

One option was passing R_MIN (len, MAX_INSTRUCTION_SIZE) instead of len, when creating the Python buffer. However, this has 2 problems: MAX_INSTRUCTION_SIZE is architecture-specific. We could retrieve it from the Python plugin on instantiation, but RAnalPlugin does not contain a proper field for storing it (instead, relying on archinfo to be used for retrieving this information). Calling archinfo every time to retrieve the maximum instruction size doesn't look right to me, and neither does storing the value statically in a global variable.

Instead, I rewrote the buffer passing code to use Python's buffer protocol and memory views - the recommended way of passing binary data to Python from C. By using the buffer protocol, we describe the buffer to Python (via a pointer to the buffer and a length), and Python looks directly at the memory provided, without copying it anywhere.

Caveat: the analysis and disassembly Python functions now receive a memoryview object instead of a bytes object. While it can be sliced as a regular array, it has some limitations. A bytes object based on it can be obtained by calling its tobytes method (which will initiate the copy - but can be called on a small slice, to save time and memory).

---

Additionally, and unrelated to the subject of this PR, I added a couple of missing refcount decrements to the analysis code.

[aronsky]

One more thing. I tried calling Py_DECREF on the memview variable after using it, but got segfaults. I am aware that I'm passing it with N instead of O, which doesn't increment the refcount - when passing with O, callling Py_DECREF afterwards works, but I don't see the difference between that and simply passing it with the N format string.

[XVilka]

Thank you, this looks good. Please update also the radare2book https://github.com/radare/radare2book/tree/master/plugins

[XVilka]

If I might ask, what Python version do you use? Python 2 or 3?

[aronsky]

Hi, I totally forgot to test it on Python 2 (I'm using 3 usually) - though I did make sure to use functions available in both 2 and 3 (Memory views have been added to Python in version 2.7).

However, checking now, it fails to build Python 2 anyway, unrelated to my changes (due to the use of PyUnicode_AsUTF8AndSize in py_io_read).

[XVilka]

Yes, and we plan to drop Python 2 support totally.

[XVilka]

See https://github.com/radare/radare2-bindings/issues/206 , so if you want you can remove Python 2 quirks in another PR.