Build by default with optimization level -O2 for GCC

before -O0 was used by default, that caused pretty bad performance.

Amd64, i6700k, 4200MHz, GNU/Linux (non-RT kernel), gcc 7.2.0

-------------------------------------
Optimization | EN/ENO |no EN/ENO |
-------------------------------------
default | 11 | 9.5 |
-O3 | 3.9 | 5.2 |
-O2 | 4 | 4.8 |
-Os | 4.1 | 3.5 |
-Ofast | 3.9 | 5.2 |
-------------------------------------

ARM, BBB Cortex-A8, 600Mhz, GNU/Linux, gcc 4.6.3

-------------------------------------
Optimization | EN/ENO |no EN/ENO |
-------------------------------------
default | 273 | 226 |
-O3 | 141.8 | 106.2 |
-O2 | 142 | 107 |
-Os | 152.5 | 112.2 |
-Ofast | 141.7 | 106.2 |
-------------------------------------

For embedded systems with size constaints (like Cortex-Mx, AVR and so
on) I usually use -Os. It gets pretty good results. For
GNU/Linux-based systems -O2 is usually a good choice, as you see the
test results.

#!/usr/bin/env python
# -*- coding: utf-8 -*-
#
# See COPYING.Runtime file for copyrights details.
#

from __future__ import absolute_import
from ctypes import CDLL, RTLD_GLOBAL, pointer, c_int, POINTER, c_char, create_string_buffer


def TryPreloadXenomai():
    """
    Xenomai 3 (at least for version <= 3.0.6) do not handle properly dlclose
    of shared objects whose dlopen did trigger xenomai_init.
    As a workaround, this pre-loads xenomai libraries that need to be
    initialized and call xenomai_init once for all.

    Xenomai auto init of libs MUST be disabled (see --auto-init-solib in xeno-config)
    """
    try:
        for name in ["cobalt", "modechk", "copperplate", "alchemy"]:
            globals()[name] = CDLL("lib"+name+".so", mode=RTLD_GLOBAL)
        cobalt.xenomai_init(pointer(c_int(0)), pointer((POINTER(c_char)*2)(create_string_buffer("prog_name"), None)))
    except Exception:
        pass