610   felem ftmp, ftmp2;  in felem_inv()  local
615   felem_square(ftmp, in); /* 2^1 */  in felem_inv()
616   felem_mul(ftmp, in, ftmp); /* 2^2 - 2^0 */  in felem_inv()
617   felem_assign(e2, ftmp);  in felem_inv()
618   felem_square(ftmp, ftmp); /* 2^3 - 2^1 */  in felem_inv()
619   felem_square(ftmp, ftmp); /* 2^4 - 2^2 */  in felem_inv()
620   felem_mul(ftmp, ftmp, e2); /* 2^4 - 2^0 */  in felem_inv()
621   felem_assign(e4, ftmp);  in felem_inv()
622   felem_square(ftmp, ftmp); /* 2^5 - 2^1 */  in felem_inv()
623   felem_square(ftmp, ftmp); /* 2^6 - 2^2 */  in felem_inv()
624   felem_square(ftmp, ftmp); /* 2^7 - 2^3 */  in felem_inv()
625   felem_square(ftmp, ftmp); /* 2^8 - 2^4 */  in felem_inv()
626   felem_mul(ftmp, ftmp, e4); /* 2^8 - 2^0 */  in felem_inv()
627   felem_assign(e8, ftmp);  in felem_inv()
629     felem_square(ftmp, ftmp);  in felem_inv()
631   felem_mul(ftmp, ftmp, e8); /* 2^16 - 2^0 */  in felem_inv()
632   felem_assign(e16, ftmp);  in felem_inv()
634     felem_square(ftmp, ftmp);  in felem_inv()
636   felem_mul(ftmp, ftmp, e16); /* 2^32 - 2^0 */  in felem_inv()
637   felem_assign(e32, ftmp);  in felem_inv()
639     felem_square(ftmp, ftmp);  in felem_inv()
641   felem_assign(e64, ftmp);  in felem_inv()
642   felem_mul(ftmp, ftmp, in); /* 2^64 - 2^32 + 2^0 */  in felem_inv()
644     felem_square(ftmp, ftmp);  in felem_inv()
667   felem_mul(out, ftmp2, ftmp); /* 2^256 - 2^224 + 2^192 + 2^96 - 3 */  in felem_inv()