According to issue 9203 the computed gotos should be enabled by default since 3.2.
However, it is not visible from the interpreter.

Python 3.2.2 (default, Sep  7 2011, 10:55:43) 
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
>>> from sysconfig import get_config_var
>>> get_config_var('HAVE_COMPUTED_GOTOS')
1
>>> get_config_var('USE_COMPUTED_GOTOS')
0

> According to issue 9203 the computed gotos should be enabled by default since 3.2.
> However, it is not visible from the interpreter.
> 
> Python 3.2.2 (default, Sep  7 2011, 10:55:43) 
> [GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
> >>> from sysconfig import get_config_var
> >>> get_config_var('HAVE_COMPUTED_GOTOS')
> 1
> >>> get_config_var('USE_COMPUTED_GOTOS')
> 0

This looks like the system Python compiled by Apple. What about a
hand-compiled Python?

Just compiled 3.2 on Debian:

>>> sysconfig.get_config_var('USE_COMPUTED_GOTOS')
0

> This looks like the system Python compiled by Apple. What about a
> hand-compiled Python?

Forget this. I get the same results here.

However, if you add a "#error" at the right place in ceval.c, you'll see that computed gotos are enabled. It seems more of a sysconfig bug or limitation. In pyconfig.h, I have the following:

$ \grep GOTO pyconfig.h
#define HAVE_COMPUTED_GOTOS 1
/* #undef USE_COMPUTED_GOTOS */

Which apparently sysconfig translates as USE_COMPUTED_GOTOS being equal to 0, which I think is wrong.

Extract of configure.in:
-------------------------------------
# Check for --with-computed-gotos
AC_MSG_CHECKING(for --with-computed-gotos)
AC_ARG_WITH(computed-gotos,
            AS_HELP_STRING([--with(out)-computed-gotos],
                           [Use computed gotos in evaluation loop (enabled by default on supported compilers)]),
[
if test "$withval" = yes
then 
  AC_DEFINE(USE_COMPUTED_GOTOS, 1,
  [Define if you want to use computed gotos in ceval.c.]) 
  AC_MSG_RESULT(yes)
fi
if test "$withval" = no
then 
  AC_DEFINE(USE_COMPUTED_GOTOS, 0,
  [Define if you want to use computed gotos in ceval.c.]) 
  AC_MSG_RESULT(no)
fi
],
[AC_MSG_RESULT(no value specified)])
-------------------------------------

Extract of my config.log:
-------------------------------------
configure:13788: checking whether gcc -pthread supports computed gotos
configure:13811: gcc -pthread -o conftest    conftest.c -lpthread -ldl  -lutil >&5
configure:13811: $? = 0
configure:13811: ./conftest
configure:13811: $? = 0
configure:13822: result: yes
configure:13832: checking for --with-computed-gotos
configure:13856: result: no value specified
-------------------------------------

Extract of my pyconfig.h
-------------------------------------
/* Define if you want to use computed gotos in ceval.c. */
/* #undef USE_COMPUTED_GOTOS */
-------------------------------------

Computed goto are disabled (in my Python 3.3 on Linux with GCC), I checked by adding #error to ceval.c.

> Extract of my pyconfig.h
> -------------------------------------
> /* Define if you want to use computed gotos in ceval.c. */
> /* #undef USE_COMPUTED_GOTOS */
> -------------------------------------

Ok, now read ceval.c:

#ifdef HAVE_COMPUTED_GOTOS
    #ifndef USE_COMPUTED_GOTOS
    #define USE_COMPUTED_GOTOS 1
    #endif
#else
    #if defined(USE_COMPUTED_GOTOS) && USE_COMPUTED_GOTOS
    #error "Computed gotos are not supported on this compiler."
    #endif
    #undef USE_COMPUTED_GOTOS
    #define USE_COMPUTED_GOTOS 0
#endif

With the #error, I can confirm that computed gotos are enabled on OS X.

About sysconfig, we may change the code to set None if the value is "undef". I don't know the impact.


--- a/Lib/sysconfig.py
+++ b/Lib/sysconfig.py
@@ -404,7 +404,7 @@
         else:
             m = undef_rx.match(line)
             if m:
-                vars[m.group(1)] = 0
+                vars[m.group(1)] = None
     return vars

> With the #error, I can confirm that computed gotos are enabled on OS X.
> 
> About sysconfig, we may change the code to set None if the value is "undef". I don't know the impact.

Actually, I think sysconfig does the right thing when it comes to all
the configure-generated HAVE_XXX variables (due to the peculiar way the
configure/pyconfig.h couple works).

It's just that USE_COMPUTED_GOTOS is a ternary variable (0/1/undefined),
so maybe we should accept as a limitation that sysconfig returns a
misleading result here.

> Computed goto are disabled (in my Python 3.3 on Linux with GCC),
> I checked by adding #error to ceval.c.

Hum, I don't understand how, but I missed the #error failure. Computed goto *are enabled* by default.

> Actually, I think sysconfig does the right thing when it comes to all
> the configure-generated HAVE_XXX variables

I agree.  The 'configure' script could be improved, though.  If we truly
want to enable this by default, then the defaulting should be moved to
configure.  This will give a more accurate portrayal in sysconfig.
Also, we can do the USE but !HAVE check in configure, which is better
anyway because the error is caught earlier.

I am thinking something like the attached.  Thoughts?

P.S. We could probably get rid of the HAVE macro all together by doing
all the work in the 'configure' script.

> If we truly want to enable this by default, then the defaulting should be moved to
> configure.  This will give a more accurate portrayal in sysconfig.
This sounds good.  (I know little about configure/pyconfig.h, but making sysconfig more accurate is a valuable result.)

> P.S. We could probably get rid of the HAVE macro all together by doing
> all the work in the 'configure' script.
Would that be a breach of backward compatibility for sysconfig?

>> P.S. We could probably get rid of the HAVE macro all together by doing
>> all the work in the 'configure' script.
> Would that be a breach of backward compatibility for sysconfig?

Yeah, I think so.  Best to leave it alone then.