The latter is 1 cycle faster on a cortex-53 and since the operands are
bytewise (or larger) bitmask (impossible to overflow to zero) both are
equivalent.