The patch unifies float("inf") and repr(float("inf")) on all platforms.

Allows dictionaries to be pre-sized (upto 255 elements) saving time lost
to re-sizes with their attendant mallocs and re-insertions.

Has zero effect on small dictionaries (5 elements or fewer), a slight
benefit for dicts upto 22 elements (because they had to resize once
anyway), and more benefit for dicts upto 255 elements (saving multiple
resizes during the build-up and reducing the number of collisions on
the first insertions).  Beyond 255 elements, there is no addional benefit.

New opcode, STORE_MAP saves the compiler from awkward stack manipulations
and specializes for dicts using PyDict_SetItem instead of PyObject_SetItem.

Old disassembly:
              0 BUILD_MAP                0
              3 DUP_TOP
              4 LOAD_CONST               1 (1)
              7 ROT_TWO
              8 LOAD_CONST               2 ('x')
             11 STORE_SUBSCR
             12 DUP_TOP
             13 LOAD_CONST               3 (2)
             16 ROT_TWO
             17 LOAD_CONST               4 ('y')
             20 STORE_SUBSCR

New disassembly:
              0 BUILD_MAP                0
              3 LOAD_CONST               1 (1)
              6 LOAD_CONST               2 ('x')
              9 STORE_MAP
             10 LOAD_CONST               3 (2)
             13 LOAD_CONST               4 ('y')
             16 STORE_MAP

Amaury's ideas works great. Should we build the Python core with WINVER=0x0500 and _WIN32_WINNT=0x0500, too?

is now simpler.  Thanks Mark Dickinson.

I had to modify Brett's patch slightly.

pointer copy and object increment in one pass.  For small lists,
save the overhead of the call to memcpy() -- this comes up in
calls like f(*listcomp).

The detection now works on Unix with Makefile, Makefile with VPATH and on Windows.

Thanks to Ian Bruntlett and Robert Lehmann.