static void fesquare(uint32_t out[9], const uint32_t in1[9]) {
  { const uint32_t x15 = in1[8];
  { const uint32_t x16 = in1[7];
  { const uint32_t x14 = in1[6];
  { const uint32_t x12 = in1[5];
  { const uint32_t x10 = in1[4];
  { const uint32_t x8 = in1[3];
  { const uint32_t x6 = in1[2];
  { const uint32_t x4 = in1[1];
  { const uint32_t x2 = in1[0];
  { ℤ x17 = ((((uint64_t)x2 * x15) + ((0x2 * ((uint64_t)x4 * x16)) + (((uint64_t)x6 * x14) + (((uint64_t)x8 * x12) + ((0x2 * ((uint64_t)x10 * x10)) + (((uint64_t)x12 * x8) + (((uint64_t)x14 * x6) + ((0x2 * ((uint64_t)x16 * x4)) + ((uint64_t)x15 * x2))))))))) +ℤ ((((uint64_t)x14 * x15) + ((0x2 * ((uint64_t)x16 * x16)) + ((uint64_t)x15 * x14))) +ℤ ((0x200000 *ℤ ((0x2 * ((uint64_t)x16 * x15)) + (0x2 * ((uint64_t)x15 * x16)))) +ℤ (0x80000000000 *ℤ ((uint64_t)x15 * x15)))));
  { uint64_t x18 = ((((uint64_t)x2 * x16) + (((uint64_t)x4 * x14) + (((uint64_t)x6 * x12) + (((uint64_t)x8 * x10) + (((uint64_t)x10 * x8) + (((uint64_t)x12 * x6) + (((uint64_t)x14 * x4) + ((uint64_t)x16 * x2)))))))) + (((uint64_t)x15 * x15) + (((uint64_t)x12 * x15) + (((uint64_t)x14 * x16) + (((uint64_t)x16 * x14) + ((uint64_t)x15 * x12))))));
  { uint64_t x19 = ((((uint64_t)x2 * x14) + ((0x2 * ((uint64_t)x4 * x12)) + ((0x2 * ((uint64_t)x6 * x10)) + (((uint64_t)x8 * x8) + ((0x2 * ((uint64_t)x10 * x6)) + ((0x2 * ((uint64_t)x12 * x4)) + ((uint64_t)x14 * x2))))))) + (((0x2 * ((uint64_t)x16 * x15)) + (0x2 * ((uint64_t)x15 * x16))) + ((0x2 * ((uint64_t)x10 * x15)) + ((0x2 * ((uint64_t)x12 * x16)) + (((uint64_t)x14 * x14) + ((0x2 * ((uint64_t)x16 * x12)) + (0x2 * ((uint64_t)x15 * x10))))))));
  { uint64_t x20 = ((((uint64_t)x2 * x12) + ((0x2 * ((uint64_t)x4 * x10)) + (((uint64_t)x6 * x8) + (((uint64_t)x8 * x6) + ((0x2 * ((uint64_t)x10 * x4)) + ((uint64_t)x12 * x2)))))) + ((((uint64_t)x14 * x15) + ((0x2 * ((uint64_t)x16 * x16)) + ((uint64_t)x15 * x14))) + (((uint64_t)x8 * x15) + ((0x2 * ((uint64_t)x10 * x16)) + (((uint64_t)x12 * x14) + (((uint64_t)x14 * x12) + ((0x2 * ((uint64_t)x16 * x10)) + ((uint64_t)x15 * x8))))))));
  { uint64_t x21 = ((((uint64_t)x2 * x10) + (((uint64_t)x4 * x8) + (((uint64_t)x6 * x6) + (((uint64_t)x8 * x4) + ((uint64_t)x10 * x2))))) + ((((uint64_t)x12 * x15) + (((uint64_t)x14 * x16) + (((uint64_t)x16 * x14) + ((uint64_t)x15 * x12)))) + (((uint64_t)x6 * x15) + (((uint64_t)x8 * x16) + (((uint64_t)x10 * x14) + (((uint64_t)x12 * x12) + (((uint64_t)x14 * x10) + (((uint64_t)x16 * x8) + ((uint64_t)x15 * x6)))))))));
  { uint64_t x22 = ((((uint64_t)x2 * x8) + ((0x2 * ((uint64_t)x4 * x6)) + ((0x2 * ((uint64_t)x6 * x4)) + ((uint64_t)x8 * x2)))) + (((0x2 * ((uint64_t)x10 * x15)) + ((0x2 * ((uint64_t)x12 * x16)) + (((uint64_t)x14 * x14) + ((0x2 * ((uint64_t)x16 * x12)) + (0x2 * ((uint64_t)x15 * x10)))))) + ((0x2 * ((uint64_t)x4 * x15)) + ((0x2 * ((uint64_t)x6 * x16)) + (((uint64_t)x8 * x14) + ((0x2 * ((uint64_t)x10 * x12)) + ((0x2 * ((uint64_t)x12 * x10)) + (((uint64_t)x14 * x8) + ((0x2 * ((uint64_t)x16 * x6)) + (0x2 * ((uint64_t)x15 * x4)))))))))));
  { uint64_t x23 = ((((uint64_t)x2 * x6) + ((0x2 * ((uint64_t)x4 * x4)) + ((uint64_t)x6 * x2))) + (((uint64_t)x8 * x15) + ((0x2 * ((uint64_t)x10 * x16)) + (((uint64_t)x12 * x14) + (((uint64_t)x14 * x12) + ((0x2 * ((uint64_t)x16 * x10)) + ((uint64_t)x15 * x8)))))));
  { uint64_t x24 = ((((uint64_t)x2 * x4) + ((uint64_t)x4 * x2)) + (((uint64_t)x6 * x15) + (((uint64_t)x8 * x16) + (((uint64_t)x10 * x14) + (((uint64_t)x12 * x12) + (((uint64_t)x14 * x10) + (((uint64_t)x16 * x8) + ((uint64_t)x15 * x6))))))));
  { uint64_t x25 = (((uint64_t)x2 * x2) + ((0x2 * ((uint64_t)x4 * x15)) + ((0x2 * ((uint64_t)x6 * x16)) + (((uint64_t)x8 * x14) + ((0x2 * ((uint64_t)x10 * x12)) + ((0x2 * ((uint64_t)x12 * x10)) + (((uint64_t)x14 * x8) + ((0x2 * ((uint64_t)x16 * x6)) + (0x2 * ((uint64_t)x15 * x4))))))))));
  { uint32_t x26 = (uint32_t) (x23 >> 0x15);
  { uint32_t x27 = ((uint32_t)x23 & 0x1fffff);
  { ℤ x28 = (x17 >>ℤ 0x15);
  { uint32_t x29 = (x17 & 0x1fffff);
  { ℤ x30 = ((0x200000 *ℤ x28) +ℤ x29);
  { ℤ x31 = (x30 >>ℤ 0x15);
  { uint32_t x32 = (x30 & 0x1fffff);
  { ℤ x33 = ((x26 + x22) +ℤ x31);
  { uint64_t x34 = (x33 >> 0x16);
  { uint32_t x35 = (x33 & 0x3fffff);
  { ℤ x36 = (x25 +ℤ x31);
  { uint64_t x37 = (x36 >> 0x16);
  { uint32_t x38 = (x36 & 0x3fffff);
  { uint64_t x39 = (x34 + x21);
  { uint32_t x40 = (uint32_t) (x39 >> 0x15);
  { uint32_t x41 = ((uint32_t)x39 & 0x1fffff);
  { uint64_t x42 = (x37 + x24);
  { uint32_t x43 = (uint32_t) (x42 >> 0x15);
  { uint32_t x44 = ((uint32_t)x42 & 0x1fffff);
  { uint64_t x45 = (x40 + x20);
  { uint32_t x46 = (uint32_t) (x45 >> 0x15);
  { uint32_t x47 = ((uint32_t)x45 & 0x1fffff);
  { uint32_t x48 = (x43 + x27);
  { uint32_t x49 = (x48 >> 0x15);
  { uint32_t x50 = (x48 & 0x1fffff);
  { uint64_t x51 = (x46 + x19);
  { uint32_t x52 = (uint32_t) (x51 >> 0x16);
  { uint32_t x53 = ((uint32_t)x51 & 0x3fffff);
  { uint64_t x54 = (x52 + x18);
  { uint32_t x55 = (uint32_t) (x54 >> 0x15);
  { uint32_t x56 = ((uint32_t)x54 & 0x1fffff);
  { uint32_t x57 = (x55 + x32);
  { uint32_t x58 = (x57 >> 0x15);
  { uint32_t x59 = (x57 & 0x1fffff);
  { uint32_t x60 = ((0x200000 * x58) + x59);
  { uint32_t x61 = (x60 >> 0x15);
  { uint32_t x62 = (x60 & 0x1fffff);
  { uint32_t x63 = ((x49 + x35) + x61);
  { uint32_t x64 = (x63 >> 0x16);
  { uint32_t x65 = (x63 & 0x3fffff);
  { uint32_t x66 = (x38 + x61);
  { uint32_t x67 = (x66 >> 0x16);
  { uint32_t x68 = (x66 & 0x3fffff);
  out[0] = x68;
  out[1] = (x67 + x44);
  out[2] = x50;
  out[3] = x65;
  out[4] = (x64 + x41);
  out[5] = x47;
  out[6] = x53;
  out[7] = x56;
  out[8] = x62;
  }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}
}