Francesco,

You have my +1 for implementing both 1 and 2 below.
"""
1. Use timegm(3) function where HAVE_TIMEGM is defined (i have a working patch for it)

2. Implement a more portable timegm function with tzset and mktime where HAVE_MKTIME and 
HAVE_WORKING_TZSET is defined (i have a working patch for it)
"""

I don't think "3. Doing some calculation taking calendar.timegm as example" is a good idea.  IMHO, its is more important that timegm is a proper inverse of gmtime on any given platform than to have the same values produced by timegm across platforms.  If system timegm (or mktime) thinks that 1900 is a leap year, for example, python should not attempt to correct that.  Maybe doing "some calculation" on systems that don't have mktime is a reasonable last resort, but I am not sure it is worth the effort.