/* calculate the 32-bit product of unsigned 16-bit value op16 and 32-bit value op32 */
static inline uint32_t mul_16_32(uint16_t op16, uint32_t op32) \
{
    unsigned t1, r;
    asm (
        "mulu   %[a], %[b]   \n"
        "swap.w %[b], %[b]   \n"
        "sts    macl, %[r]   \n"
        "mulu   %[a], %[b]   \n"
        "sts    macl, %[t1]  \n"
        "shll16 %[t1]        \n"
        "add    %[t1], %[r]  \n"
        : [r]  "=&r" (r),
          [t1] "=r" (t1),
        : [a]  "r"   (op16),
          [b]  "r"   (op32),
    );
    return r;
}