Here are some timings for dict creation, created with the attached script.  
They're not particularly scientific, but they at least show that this one-
line optimization can make a significant difference.

Typical results on my machine (OS X 10.5.6/Intel), 32-bit non-debug build 
of the trunk (r69442):  before

dict creation (selected):  1.95572495461
dict creation (shuffled):  1.98964595795
dict creation:  1.78589916229

and after:

dict creation (selected):  1.7055079937   # (14% speedup)
dict creation (shuffled):  1.5843398571   # (25% speedup)
dict creation:  1.32362794876             # (34% speedup)

BTW, all tests pass on my machine with this patch applied.