Test comprehensions and generator expressions with loop optimisations

[scoder]

Generator expressions and comprehensions seem to need a lot more testing, especially in combination with all the loop optimisations that we have: lists/tuples, range(), enumerate(), bytes/bytearray/unicode, memoryviews, C arrays, you name it.

At least the bytearray iteration fails to optimise inside of a generator expression, probably others.

This seems worth adding as dedicated test modules, e.g. genexpr_loop_optimisations_py.py (and possibly genexpr_loop_optimisations_cy.pyx).

Found as part of the change in #4254 (which broke memoryview iteration).

[da-woods]

To me the bytearray iteration looks OK in a generator expression

Just as a short example

def f(x: bytearray):
    count = 0
    for y in x:
        count += 1
    return (y==count for y in x)

Snippet of the C code for the generator:

__pyx_t_2 = PyByteArray_GET_SIZE(__pyx_t_1); if (unlikely(__pyx_t_2 == ((Py_ssize_t)-1))) __PYX_ERR(0, 7, __pyx_L1_error)
  for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_2; __pyx_t_4++) {
    __pyx_t_3 = __pyx_t_4;
    __pyx_cur_scope->__pyx_v_y = ((unsigned char)(PyByteArray_AS_STRING(__pyx_t_1)[__pyx_t_3]));

and for the plain loop:

__pyx_t_2 = PyByteArray_GET_SIZE(__pyx_t_1); if (unlikely(__pyx_t_2 == ((Py_ssize_t)-1))) __PYX_ERR(0, 5, __pyx_L1_error)
  for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_2; __pyx_t_4++) {
    __pyx_t_3 = __pyx_t_4;
    __pyx_v_y = ((unsigned char)(PyByteArray_AS_STRING(__pyx_t_1)[__pyx_t_3]));

So they're doing substantially the same thing - getting the size with PyByteArray_GET_SIZE and indexing into the C string from PyByteArray_AS_STRING.

Doesn't mean it shouldn't be tested of course (or that you haven't found a slightly different case where it doesn't optimize)