I may have had some old code in the build or something - I did a clean
rebuild to look at some of the slowdowns and the fastglobals_test
benchmarks are now even better than Neils. Pybench is still overall
slower but it's mostly in float operations, which seems odd and I'd
discount it unless someone else can recreate it. I've updated the
spreadsheet I linked before with the updated timings, along with my
microbenchmark and pystone results. Very impressive speedup on pystone,
no doubt because of the heavy use of globals.