This brings in ~8x speedup with tcc and ~14x speedup for gccO2 compared to what
we already had for pystone.py already compiled by Cython.

However Cython compilation time is 1.7x slower and compilation time for gcc*
also raises significantly. For tcc it is "only" 30% more for compilation, but
still the total tcc compilation time is very very small.

name                    old time/op    new time/op    delta
pycompile+pyimport        13.0ms ± 0%    13.0ms ± 0%       ~     (all equal)
pyimport                  12.0ms ± 0%    12.0ms ± 0%       ~     (all equal)
cython/py                  980ms ± 1%     982ms ± 1%       ~     (p=0.666 n=10+10)
cython/pyx                 983ms ± 1%    1692ms ± 1%    +72.13%  (p=0.000 n=9+10)
gcc/compile+link/py        603ms ± 0%     603ms ± 0%       ~     (p=0.766 n=8+10)
gcc/compile+linkO2/py      2.15s ± 0%     2.15s ± 0%       ~     (p=0.749 n=10+10)
gcc/compile+link/pyx       603ms ± 0%    1106ms ± 0%    +83.57%  (p=0.000 n=9+10)
gcc/compile+linkO2/pyx     2.15s ± 0%     3.65s ± 0%    +69.87%  (p=0.000 n=10+10)
tcc/compile+link/py       36.5ms ± 1%    36.0ms ± 0%     -1.37%  (p=0.022 n=10+9)
tcc/compile+link/pyx      36.7ms ± 2%    47.5ms ± 3%    +29.43%  (p=0.000 n=10+10)

name                    old pystone/s  new pystone/s  delta
pystone/py                  227k ± 2%      228k ± 1%       ~     (p=0.182 n=10+9)
pystone/py/gcc              220k ± 3%      221k ± 4%       ~     (p=0.796 n=10+10)
pystone/py/gccO2            433k ± 2%      431k ± 5%       ~     (p=1.000 n=10+10)
pystone/pyx/gcc             221k ± 1%     1855k ± 1%   +737.67%  (p=0.000 n=9+8)
pystone/pyx/gccO2           435k ± 1%     6239k ± 2%  +1335.52%  (p=0.000 n=10+10)
pystone/py/tcc              219k ± 2%      219k ± 0%       ~     (p=0.315 n=10+10)
pystone/pyx/tcc             222k ± 2%     1826k ± 0%   +723.78%  (p=0.000 n=9+9)

This will be type-annotated version of pystone.py, but for now it is
pristine copy, thus equal to it in all benchmarks.

We know from past benchmarks that most of the time is in compile.

Don't continue on `if $CC == tcc` -> instead put actions that are not
appropriate for `tcc -O2` under `if $CC != tcc`. We will add other
actions to that loop - that's the reason.

In preparation to add pystone_pyx.pyx

Cython is currently slow:

	name                time/op
	pycompile+pyimport  13.0ms ± 0%
	pyimport            12.0ms ± 0%
	cython/py            983ms ± 1%

We mesure time, not operations per second.

In preparation to add pystone_pyx.pyx

In preparation that in addition to pystone.pt there will be also
pystone_pyx.pyx with type annotations.

pystone can say, e.g. "this machine benchmarks at 1.85343e+06 pystones"

It is close to gcc.

Previously it was failing with master:

(neo) (z-dev) (g.env) kirr@deco:~/src/tools/py/cython/t/compilebench$ ./doit.sh
/home/kirr/src/tools/py/cython/Cython/Compiler/Main.py:344: FutureWarning: Cython directive 'language_level' not set, using '3str' for now (Py3). This has changed from earlier releases! File: /home/kirr/src/tools/py/cython/t/compilebench/pystone_pyxtest.py
  tree = Parsing.p_module(s, pxd, full_module_name)

Error compiling Cython file:
------------------------------------------------------------
...
TRUE = 1
FALSE = 0

def main(loops=LOOPS):
    benchtime, stones = pystones(loops)
    print "Pystone(%s) time for %d passes = %g" % \
         ^
------------------------------------------------------------

pystone_pyxtest.py:62:10: Syntax error in simple statement list

* master: (280 commits)
  Fix compilation in PyPy.
  Update changelog.
  Revert "Disable "c_api_binop_methods" directive in 0.29.x and provide it only as an enabled forward option."
  Disable "c_api_binop_methods" directive in 0.29.x and provide it only as an enabled forward option.
  Update changelog.
  Use Py_SET_SIZE() and Py_SET_REFCNT() on Python 3.9.0a4 and newer (GH-3639)
  Update changelog.
  Use Py_SET_SIZE() and Py_SET_REFCNT() on Python 3.9.0a4 and newer (GH-3639)
  Limited API updates and cleanup for #2056. GH-3635)
  Update changelog.
  Update changelog.
  Emit an error when typeid() is used outside of C++ mode (GH-3637)
  Fix overflow handling for abs() calls on signed integer types (GH-3634)
  Update changelog.
  Fix overflow handling for abs() calls on signed integer types (GH-3634)
  Fix handling of "no-cpp" test tag for error tests, where "cpp" is already removed from the list of test backends.
  Update changelog.
  Avoid "uninitialised" warnings for sizeof and typeid (GH-3631)
  Emit an error when typeid() is used outside of C++ mode (GH-3637)
  Fix handling of "no-cpp" test tag for error tests, where "cpp" is already removed from the list of test backends.
  ...

Revert "Disable "c_api_binop_methods" directive in 0.29.x and provide it only as an enabled forward option."

This reverts commit 4ac64ddf.

* Add __Pyx_SET_SIZE() function: use Py_SET_SIZE() on Python 3.9.0a4
and newer, or use Py_SIZE() as an l-value on older Python versions.

Py_SIZE() must not be used as an l-value anymore in Python 3.9:
Py_SET_SIZE() must be used instead:

* https://bugs.python.org/issue39573
* https://docs.python.org/dev/c-api/structures.html#c.Py_SET_SIZE

* Add __Pyx_SET_REFCNT() function: use Py_SET_REFCNT() on Python 3.9.0a4
and newer, or use Py_REFCNT() as an l-value on older Python versions.

Py_REFCNT() must not be used as an l-value anymore in Python 3.9:
Py_SET_REFCNT() must be used instead:

* https://bugs.python.org/issue39573
* https://docs.python.org/dev/c-api/structures.html#c.Py_SET_REFCNT

Use it in ModuleNode.generate_usr_dealloc_call():

* Replace ++Py_REFCNT(o) with __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1)
* Replace --Py_REFCNT(o) with __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1)

* Add __Pyx_SET_SIZE() function: use Py_SET_SIZE() on Python 3.9.0a4
and newer, or use Py_SIZE() as an l-value on older Python versions.

Py_SIZE() must not be used as an l-value anymore in Python 3.9:
Py_SET_SIZE() must be used instead:

* https://bugs.python.org/issue39573
* https://docs.python.org/dev/c-api/structures.html#c.Py_SET_SIZE

* Add __Pyx_SET_REFCNT() function: use Py_SET_REFCNT() on Python 3.9.0a4
and newer, or use Py_REFCNT() as an l-value on older Python versions.

Py_REFCNT() must not be used as an l-value anymore in Python 3.9:
Py_SET_REFCNT() must be used instead:

* https://bugs.python.org/issue39573
* https://docs.python.org/dev/c-api/structures.html#c.Py_SET_REFCNT

Use it in ModuleNode.generate_usr_dealloc_call():

* Replace ++Py_REFCNT(o) with __Pyx_SET_REFCNT(o, Py_REFCNT(o) + 1)
* Replace --Py_REFCNT(o) with __Pyx_SET_REFCNT(o, Py_REFCNT(o) - 1)

Fixes #1911

Fixes #1911

Fix handling of "no-cpp" test tag for error tests, where "cpp" is already removed from the list of test backends.

Closes #3575

Fix handling of "no-cpp" test tag for error tests, where "cpp" is already removed from the list of test backends.

Closes #3575