static void femul(uint64_t out[8], const uint64_t in1[8], const uint64_t in2[8]) {
  { const uint64_t x16 = in1[7];
  { const uint64_t x17 = in1[6];
  { const uint64_t x15 = in1[5];
  { const uint64_t x13 = in1[4];
  { const uint64_t x11 = in1[3];
  { const uint64_t x9 = in1[2];
  { const uint64_t x7 = in1[1];
  { const uint64_t x5 = in1[0];
  { const uint64_t x30 = in2[7];
  { const uint64_t x31 = in2[6];
  { const uint64_t x29 = in2[5];
  { const uint64_t x27 = in2[4];
  { const uint64_t x25 = in2[3];
  { const uint64_t x23 = in2[2];
  { const uint64_t x21 = in2[1];
  { const uint64_t x19 = in2[0];
  { uint128_t x32 = (((uint128_t)(x11 + x16) * (x25 + x30)) - ((uint128_t)x11 * x25));
  { uint128_t x33 = ((((uint128_t)(x9 + x17) * (x25 + x30)) + ((uint128_t)(x11 + x16) * (x23 + x31))) - (((uint128_t)x9 * x25) + ((uint128_t)x11 * x23)));
  { uint128_t x34 = ((((uint128_t)(x7 + x15) * (x25 + x30)) + (((uint128_t)(x9 + x17) * (x23 + x31)) + ((uint128_t)(x11 + x16) * (x21 + x29)))) - (((uint128_t)x7 * x25) + (((uint128_t)x9 * x23) + ((uint128_t)x11 * x21))));
  { uint128_t x35 = ((((uint128_t)(x5 + x13) * (x25 + x30)) + (((uint128_t)(x7 + x15) * (x23 + x31)) + (((uint128_t)(x9 + x17) * (x21 + x29)) + ((uint128_t)(x11 + x16) * (x19 + x27))))) - (((uint128_t)x5 * x25) + (((uint128_t)x7 * x23) + (((uint128_t)x9 * x21) + ((uint128_t)x11 * x19)))));
  { uint128_t x36 = ((((uint128_t)(x5 + x13) * (x23 + x31)) + (((uint128_t)(x7 + x15) * (x21 + x29)) + ((uint128_t)(x9 + x17) * (x19 + x27)))) - (((uint128_t)x5 * x23) + (((uint128_t)x7 * x21) + ((uint128_t)x9 * x19))));
  { uint128_t x37 = ((((uint128_t)(x5 + x13) * (x21 + x29)) + ((uint128_t)(x7 + x15) * (x19 + x27))) - (((uint128_t)x5 * x21) + ((uint128_t)x7 * x19)));
  { uint128_t x38 = (((uint128_t)(x5 + x13) * (x19 + x27)) - ((uint128_t)x5 * x19));
  { uint128_t x39 = (((((uint128_t)x11 * x25) + ((uint128_t)x16 * x30)) + x36) + x32);
  { uint128_t x40 = ((((((uint128_t)x9 * x25) + ((uint128_t)x11 * x23)) + (((uint128_t)x17 * x30) + ((uint128_t)x16 * x31))) + x37) + x33);
  { uint128_t x41 = ((((((uint128_t)x7 * x25) + (((uint128_t)x9 * x23) + ((uint128_t)x11 * x21))) + (((uint128_t)x15 * x30) + (((uint128_t)x17 * x31) + ((uint128_t)x16 * x29)))) + x38) + x34);
  { uint128_t x42 = ((((uint128_t)x5 * x25) + (((uint128_t)x7 * x23) + (((uint128_t)x9 * x21) + ((uint128_t)x11 * x19)))) + (((uint128_t)x13 * x30) + (((uint128_t)x15 * x31) + (((uint128_t)x17 * x29) + ((uint128_t)x16 * x27)))));
  { uint128_t x43 = (((((uint128_t)x5 * x23) + (((uint128_t)x7 * x21) + ((uint128_t)x9 * x19))) + (((uint128_t)x13 * x31) + (((uint128_t)x15 * x29) + ((uint128_t)x17 * x27)))) + x32);
  { uint128_t x44 = (((((uint128_t)x5 * x21) + ((uint128_t)x7 * x19)) + (((uint128_t)x13 * x29) + ((uint128_t)x15 * x27))) + x33);
  { uint128_t x45 = ((((uint128_t)x5 * x19) + ((uint128_t)x13 * x27)) + x34);
  { uint64_t x46 = (uint64_t) (x42 >> 0x34);
  { uint64_t x47 = ((uint64_t)x42 & 0xfffffffffffff);
  { uint64_t x48 = (uint64_t) (x35 >> 0x34);
  { uint64_t x49 = ((uint64_t)x35 & 0xfffffffffffff);
  { uint128_t x50 = (((uint128_t)0x10000000000000 * x48) + x49);
  { uint64_t x51 = (uint64_t) (x50 >> 0x34);
  { uint64_t x52 = ((uint64_t)x50 & 0xfffffffffffff);
  { uint128_t x53 = ((x46 + x41) + x51);
  { uint64_t x54 = (uint64_t) (x53 >> 0x34);
  { uint64_t x55 = ((uint64_t)x53 & 0xfffffffffffff);
  { uint128_t x56 = (x45 + x51);
  { uint64_t x57 = (uint64_t) (x56 >> 0x34);
  { uint64_t x58 = ((uint64_t)x56 & 0xfffffffffffff);
  { uint128_t x59 = (x54 + x40);
  { uint64_t x60 = (uint64_t) (x59 >> 0x34);
  { uint64_t x61 = ((uint64_t)x59 & 0xfffffffffffff);
  { uint128_t x62 = (x57 + x44);
  { uint64_t x63 = (uint64_t) (x62 >> 0x34);
  { uint64_t x64 = ((uint64_t)x62 & 0xfffffffffffff);
  { uint128_t x65 = (x60 + x39);
  { uint64_t x66 = (uint64_t) (x65 >> 0x34);
  { uint64_t x67 = ((uint64_t)x65 & 0xfffffffffffff);
  { uint128_t x68 = (x63 + x43);
  { uint64_t x69 = (uint64_t) (x68 >> 0x34);
  { uint64_t x70 = ((uint64_t)x68 & 0xfffffffffffff);
  { uint64_t x71 = (x66 + x52);
  { uint64_t x72 = (x71 >> 0x34);
  { uint64_t x73 = (x71 & 0xfffffffffffff);
  { uint64_t x74 = (x69 + x47);
  { uint64_t x75 = (x74 >> 0x34);
  { uint64_t x76 = (x74 & 0xfffffffffffff);
  { uint64_t x77 = ((0x10000000000000 * x72) + x73);
  { uint64_t x78 = (x77 >> 0x34);
  { uint64_t x79 = (x77 & 0xfffffffffffff);
  { uint64_t x80 = ((x75 + x55) + x78);
  { uint64_t x81 = (x80 >> 0x34);
  { uint64_t x82 = (x80 & 0xfffffffffffff);
  { uint64_t x83 = (x58 + x78);
  { uint64_t x84 = (x83 >> 0x34);
  { uint64_t x85 = (x83 & 0xfffffffffffff);
  out[0] = x85;
  out[1] = (x84 + x64);
  out[2] = x70;
  out[3] = x76;
  out[4] = x82;
  out[5] = (x81 + x61);
  out[6] = x67;
  out[7] = x79;
  }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}
}