On platforms that have compiler support for doubles, such as the stm32, there are a lot of unintended float-to-double conversions in the code. I was able to reduce move() duration from 14us to 6us by eliminating just one of these unintended casts, and loopFOC() from 144us to 50us by replacing several. Almost all files have unintended promotions in them.

These instances can be easily found with -Wdouble-promotion and -Werror=double-promotion, there are a few hundred of them.