We could keep version.split’s result and use that instead of splitting again.

I tried that and it causes a minor slow down of simple version construction.

Interesting, do you know why? Is it because it keeps around the reference for a bit longer?

Not a blocker by the way

I can only speculate here, but yeah, it seems like storing a value and then accessing it has a measurable performance impact at the level of optimization we're at now.

It still gets stored, though - the only difference should be that the ref lives longer (and gets bumped one more pair of times, probably). Unless it can reduce this (not impossible, but unlikely - numpy can do this sort of thing on unix but not windows).

I wonder if del'ing the variable afterwords would affect it.