Absolutely. Global variables are always going to need some sort of arbitration mechanism, and this is doing it in the innermost loop. Even if there's only one thread, something needs to find that out. I'm surprised it's only half the speed.

No, with sum moved into the function I get 6s for --enable-gil and 15.5s for --disable-gil on my machine, which is in the same order.

As an added bonus, the stock Debian Python 3.7 is faster than both versions.

So the alleged "faster python" promises are not kept, and neither are the "at most 15% slowdown" promises for the no-gil project.

Would be worthwhile to run something more significant, maybe pybench.

This benchmark shows abysmal numbers, no need to run a bloated test suite that is so large and convoluted that no one understands what is being measured.

A test that shows a worst possible scenario is not a benchmark, not an accurate reflection of what users will experience.