WIP: Simple Cython implementation

[matusvalo]

This PR is an alternative to #258.

It consists of 3 parts:

1. Use __slots__ in GenericContent, since should be more lightweight than regular class
2. Removes and simplifies parameters of functions in serialization.py - this is prerequisite for cythonizing + this should be faster in cpython 3.7+ [1] [2]
3. simple cythonization of serialization.py hot paths using pure python mode - this enables two possible ways of using library a) cythonized b) not cythonized

The advantage of this approach is to have one single code base which is already well tested and tuned by users. The disadvantage is that not everything can be converted to "pure" C.

What is pending:

1. performance tests (should be faster but no tests are done)
2. There is one single unittests failing (all other are passed) - this needs to be decided if the test is ignored on cython or not
3. setup.py needs to be refactored to somehow allow user to choose if he wants use cythonized version or pure python
4. CI must be set to execute tests cythoned and pure python version of library
5. I have even more aggressive cythonization of serialization.py but it is partialy broken and needs some small adjustments of serialization.py

[1] https://bugs.python.org/issue26110
[2] https://docs.python.org/3.7/whatsnew/3.7.html#cpython-bytecode-changes

[matusvalo]

There is only one difference of behavior between cythonized and pure python version of library (as long as I know) - and this is causing the only one unittest failing:

cythonized version of library due translating python function to C function is not raising FrameSyntaxError on some places - e.g.:

py-amqp/amqp/serialization.py

Lines 141 to 144 in ed75968

	raise FrameSyntaxError(
	'Unknown value in table: {0!r} ({1!r})'.format(
	ftype, type(ftype)))
	return val, offset

In cython it is tranlated to:

+141:         raise FrameSyntaxError(
  /*else*/ {
    __Pyx_GetModuleGlobalName(__pyx_t_2, __pyx_n_s_FrameSyntaxError); if (unlikely(!__pyx_t_2)) __PYX_ERR(0, 141, __pyx_L1_error)
    __Pyx_GOTREF(__pyx_t_2);
+142:             'Unknown value in table: {0!r} ({1!r})'.format(
+143:                 ftype, type(ftype)))
+144:     return val, offset

But this should not happen when broker is correctly implemented - but in any case cythonized version should not be used with nonstandard/experimental brokers.

[matusvalo]

Here is cythonization report of serialization.py:

serialization.zip

Here you can see which lines (less yellow, more pure C used) uses pure C (these yields biggest speed improvement). But this can be improved more when we will convert and statically type buffer and format of some functions - e.g.:

py-amqp/amqp/serialization.py

Line 37 in 7878f1d

def _read_item(buf, offset):

but as I have said needs more changes in library.

[auvipy]

the approach and codes look less redical than rust version.

[thedrow]

I still think we should avoid C extensions and unsafe code.
Cython code cannot be shared with CFFI code. The Rust code can be.

[auvipy]

then we could try this in another fork. celery should be python3 only

[thedrow]

I don't think you understand what I mean.
C extensions are not the right way to go.
Rust is when it is optional.

[auvipy]

https://twitter.com/pwang/status/1203336880886075392/photo/1

[matusvalo]

Also aiohttp is using cython - see:

https://github.com/aio-libs/aiohttp/tree/master/aiohttp

[matusvalo]

Closed in favor of #311