such a great post full of gems like this: "… Tanja Lange pointed me to this study by Wouter de Groot, who performed actual measures on some Cortex-M3 devices,

such a great post full of gems like this: "… Tanja Lange pointed me to this study by Wouter de Groot, who performed actual measures on some Cortex-M3 devices, from which it appeared that a short cycle count could occur not only in the documented case (operands fit in 16 bits) but also when one or both of the operands is zero or a power of 2. In the case of MUL31(), this means that the underlying umull will complete in 5 cycles, except if one or both of the two operands are zero, in which case the umull will complete in 4 cycles. …"