i still don't know this STM32H735 ARM processor manages to take 225 cycles to compute a single iteration of this loop.

the call to polyBlep() isn't a significant part of the cost because removing it doesn't dramatically impact the performance.