Neal, t and stream aren't likely to have much effect since they're used
so little. next_instr and stack_pointer are used to communicate between
loops, so they can't move inside. I eagerly await your benchmark runs. :)

Skip, I managed to reproduce the warnings with gcc-4.2.1, and while
they're wrong, I see why they're happening. The if (throwflag) block
skips to on_error, which misses the initializations of x and err. The
right way to fix it is either to eliminate the error-reporting behavior
of x and err or to extract on_error into a function, but for this patch
I would probably just keep x and err out of the loop.

Your numbers made me look closer at my results for individual tests, and
I'm now confused about how reliable pybench is. My gcc-4.0 was build
#5367 on OS X 10.4.11, and MacPorts' gcc-4.2.1 (with a necessary
configure.in tweak) still shows a 1-2% gain overall. But contrary to
your numbers, it gives me a 10% speedup in Recursion and a 1% slowdown
in CompareFloatsIntegers. My big losers are SimpleListManipulation with
an 18% loss on 4.2 and CompareInternedStrings with a 20% loss on 4.0,
but those are both small winners (~5%) on the opposite compiler! I
wouldn't be surprised if the overall numbers were different between even
slightly different machines and compilers, but I'm really surprised to
see the same tests affected in opposite directions. Is that common with
pybench and compiler changes?