Currently, Python switches from %f to %g formatting at 1e50.  This 
applies both to the old-style percent formatting, and to the newer PEP 
3101-style formatting:

>>> '%f' % 2.**166
'93536104789177786765035829293842113257979682750464.000000'
>>> '%f' % 2.**167
'1.87072e+50'

>>> format(2.**166, 'f')
'93536104789177786765035829293842113257979682750464.000000'
>>> format(2.**167, 'f')
'1.87072e+50'

The main reason for the switch seems to have been implementation 
convenience:  it makes it possible to use a fixed-size buffer in
the float formatting routines.

I propose removing this feature for Python 3.1, but leaving it in place 
for 2.7.

See

http://mail.python.org/pipermail/python-dev/2009-April/089030.html

for additional discussion.

Here's a patch which doesn't fix the '%f' -> '%g' switch, but paves the
way for that switch by getting rid of the fixed-size buffers in
formatfloat in unicodeobject.c, and in the fallback version of
PyOS_double_to_string.

No new test failures on Linux or OS X, with or without the fallback code.  
I've committed this in r72161.

The patch looks good to me.

Here's the patch that actually removes the %f -> %g switch.

Fixed in r72165.