static void femul(uint64_t out[4], const uint64_t in1[4], const uint64_t in2[4]) {
  { const uint64_t x8 = in1[3];
  { const uint64_t x9 = in1[2];
  { const uint64_t x7 = in1[1];
  { const uint64_t x5 = in1[0];
  { const uint64_t x14 = in2[3];
  { const uint64_t x15 = in2[2];
  { const uint64_t x13 = in2[1];
  { const uint64_t x11 = in2[0];
  { uint64_t x18;  uint64_t x17 = _mulx_u64(x5, x11, &x18);
  { uint64_t x21;  uint64_t x20 = _mulx_u64(x5, x13, &x21);
  { uint64_t x24;  uint64_t x23 = _mulx_u64(x5, x15, &x24);
  { uint64_t x27;  uint64_t x26 = _mulx_u64(x5, x14, &x27);
  { uint64_t x29; uint8_t x30 = _addcarryx_u64(0x0, x18, x20, &x29);
  { uint64_t x32; uint8_t x33 = _addcarryx_u64(x30, x21, x23, &x32);
  { uint64_t x35; uint8_t x36 = _addcarryx_u64(x33, x24, x26, &x35);
  { uint64_t x38; uint8_t _ = _addcarryx_u64(0x0, x36, x27, &x38);
  { uint64_t _;  uint64_t x41 = _mulx_u64(x17, 0xf83e0f83e0f83e1, &_);
  { uint64_t x45;  uint64_t x44 = _mulx_u64(x41, 0xffffffffffffffdfL, &x45);
  { uint64_t x48;  uint64_t x47 = _mulx_u64(x41, 0xffffffffffffffffL, &x48);
  { uint64_t x51;  uint64_t x50 = _mulx_u64(x41, 0xffffffffffffffffL, &x51);
  { uint8_t x54;  uint64_t x53 = _mulx_u64_out_u8(x41, 0x3, &x54);
  { uint64_t x56; uint8_t x57 = _addcarryx_u64(0x0, x45, x47, &x56);
  { uint64_t x59; uint8_t x60 = _addcarryx_u64(x57, x48, x50, &x59);
  { uint64_t x62; uint8_t x63 = _addcarryx_u64(x60, x51, x53, &x62);
  { uint8_t x64 = (x63 + x54);
  { uint64_t _; uint8_t x67 = _addcarryx_u64(0x0, x17, x44, &_);
  { uint64_t x69; uint8_t x70 = _addcarryx_u64(x67, x29, x56, &x69);
  { uint64_t x72; uint8_t x73 = _addcarryx_u64(x70, x32, x59, &x72);
  { uint64_t x75; uint8_t x76 = _addcarryx_u64(x73, x35, x62, &x75);
  { uint64_t x78; uint8_t x79 = _addcarryx_u64(x76, x38, x64, &x78);
  { uint64_t x82;  uint64_t x81 = _mulx_u64(x7, x11, &x82);
  { uint64_t x85;  uint64_t x84 = _mulx_u64(x7, x13, &x85);
  { uint64_t x88;  uint64_t x87 = _mulx_u64(x7, x15, &x88);
  { uint64_t x91;  uint64_t x90 = _mulx_u64(x7, x14, &x91);
  { uint64_t x93; uint8_t x94 = _addcarryx_u64(0x0, x82, x84, &x93);
  { uint64_t x96; uint8_t x97 = _addcarryx_u64(x94, x85, x87, &x96);
  { uint64_t x99; uint8_t x100 = _addcarryx_u64(x97, x88, x90, &x99);
  { uint64_t x102; uint8_t _ = _addcarryx_u64(0x0, x100, x91, &x102);
  { uint64_t x105; uint8_t x106 = _addcarryx_u64(0x0, x69, x81, &x105);
  { uint64_t x108; uint8_t x109 = _addcarryx_u64(x106, x72, x93, &x108);
  { uint64_t x111; uint8_t x112 = _addcarryx_u64(x109, x75, x96, &x111);
  { uint64_t x114; uint8_t x115 = _addcarryx_u64(x112, x78, x99, &x114);
  { uint64_t x117; uint8_t x118 = _addcarryx_u64(x115, x79, x102, &x117);
  { uint64_t _;  uint64_t x120 = _mulx_u64(x105, 0xf83e0f83e0f83e1, &_);
  { uint64_t x124;  uint64_t x123 = _mulx_u64(x120, 0xffffffffffffffdfL, &x124);
  { uint64_t x127;  uint64_t x126 = _mulx_u64(x120, 0xffffffffffffffffL, &x127);
  { uint64_t x130;  uint64_t x129 = _mulx_u64(x120, 0xffffffffffffffffL, &x130);
  { uint8_t x133;  uint64_t x132 = _mulx_u64_out_u8(x120, 0x3, &x133);
  { uint64_t x135; uint8_t x136 = _addcarryx_u64(0x0, x124, x126, &x135);
  { uint64_t x138; uint8_t x139 = _addcarryx_u64(x136, x127, x129, &x138);
  { uint64_t x141; uint8_t x142 = _addcarryx_u64(x139, x130, x132, &x141);
  { uint8_t x143 = (x142 + x133);
  { uint64_t _; uint8_t x146 = _addcarryx_u64(0x0, x105, x123, &_);
  { uint64_t x148; uint8_t x149 = _addcarryx_u64(x146, x108, x135, &x148);
  { uint64_t x151; uint8_t x152 = _addcarryx_u64(x149, x111, x138, &x151);
  { uint64_t x154; uint8_t x155 = _addcarryx_u64(x152, x114, x141, &x154);
  { uint64_t x157; uint8_t x158 = _addcarryx_u64(x155, x117, x143, &x157);
  { uint8_t x159 = (x158 + x118);
  { uint64_t x162;  uint64_t x161 = _mulx_u64(x9, x11, &x162);
  { uint64_t x165;  uint64_t x164 = _mulx_u64(x9, x13, &x165);
  { uint64_t x168;  uint64_t x167 = _mulx_u64(x9, x15, &x168);
  { uint64_t x171;  uint64_t x170 = _mulx_u64(x9, x14, &x171);
  { uint64_t x173; uint8_t x174 = _addcarryx_u64(0x0, x162, x164, &x173);
  { uint64_t x176; uint8_t x177 = _addcarryx_u64(x174, x165, x167, &x176);
  { uint64_t x179; uint8_t x180 = _addcarryx_u64(x177, x168, x170, &x179);
  { uint64_t x182; uint8_t _ = _addcarryx_u64(0x0, x180, x171, &x182);
  { uint64_t x185; uint8_t x186 = _addcarryx_u64(0x0, x148, x161, &x185);
  { uint64_t x188; uint8_t x189 = _addcarryx_u64(x186, x151, x173, &x188);
  { uint64_t x191; uint8_t x192 = _addcarryx_u64(x189, x154, x176, &x191);
  { uint64_t x194; uint8_t x195 = _addcarryx_u64(x192, x157, x179, &x194);
  { uint64_t x197; uint8_t x198 = _addcarryx_u64(x195, x159, x182, &x197);
  { uint64_t _;  uint64_t x200 = _mulx_u64(x185, 0xf83e0f83e0f83e1, &_);
  { uint64_t x204;  uint64_t x203 = _mulx_u64(x200, 0xffffffffffffffdfL, &x204);
  { uint64_t x207;  uint64_t x206 = _mulx_u64(x200, 0xffffffffffffffffL, &x207);
  { uint64_t x210;  uint64_t x209 = _mulx_u64(x200, 0xffffffffffffffffL, &x210);
  { uint8_t x213;  uint64_t x212 = _mulx_u64_out_u8(x200, 0x3, &x213);
  { uint64_t x215; uint8_t x216 = _addcarryx_u64(0x0, x204, x206, &x215);
  { uint64_t x218; uint8_t x219 = _addcarryx_u64(x216, x207, x209, &x218);
  { uint64_t x221; uint8_t x222 = _addcarryx_u64(x219, x210, x212, &x221);
  { uint8_t x223 = (x222 + x213);
  { uint64_t _; uint8_t x226 = _addcarryx_u64(0x0, x185, x203, &_);
  { uint64_t x228; uint8_t x229 = _addcarryx_u64(x226, x188, x215, &x228);
  { uint64_t x231; uint8_t x232 = _addcarryx_u64(x229, x191, x218, &x231);
  { uint64_t x234; uint8_t x235 = _addcarryx_u64(x232, x194, x221, &x234);
  { uint64_t x237; uint8_t x238 = _addcarryx_u64(x235, x197, x223, &x237);
  { uint8_t x239 = (x238 + x198);
  { uint64_t x242;  uint64_t x241 = _mulx_u64(x8, x11, &x242);
  { uint64_t x245;  uint64_t x244 = _mulx_u64(x8, x13, &x245);
  { uint64_t x248;  uint64_t x247 = _mulx_u64(x8, x15, &x248);
  { uint64_t x251;  uint64_t x250 = _mulx_u64(x8, x14, &x251);
  { uint64_t x253; uint8_t x254 = _addcarryx_u64(0x0, x242, x244, &x253);
  { uint64_t x256; uint8_t x257 = _addcarryx_u64(x254, x245, x247, &x256);
  { uint64_t x259; uint8_t x260 = _addcarryx_u64(x257, x248, x250, &x259);
  { uint64_t x262; uint8_t _ = _addcarryx_u64(0x0, x260, x251, &x262);
  { uint64_t x265; uint8_t x266 = _addcarryx_u64(0x0, x228, x241, &x265);
  { uint64_t x268; uint8_t x269 = _addcarryx_u64(x266, x231, x253, &x268);
  { uint64_t x271; uint8_t x272 = _addcarryx_u64(x269, x234, x256, &x271);
  { uint64_t x274; uint8_t x275 = _addcarryx_u64(x272, x237, x259, &x274);
  { uint64_t x277; uint8_t x278 = _addcarryx_u64(x275, x239, x262, &x277);
  { uint64_t _;  uint64_t x280 = _mulx_u64(x265, 0xf83e0f83e0f83e1, &_);
  { uint64_t x284;  uint64_t x283 = _mulx_u64(x280, 0xffffffffffffffdfL, &x284);
  { uint64_t x287;  uint64_t x286 = _mulx_u64(x280, 0xffffffffffffffffL, &x287);
  { uint64_t x290;  uint64_t x289 = _mulx_u64(x280, 0xffffffffffffffffL, &x290);
  { uint8_t x293;  uint64_t x292 = _mulx_u64_out_u8(x280, 0x3, &x293);
  { uint64_t x295; uint8_t x296 = _addcarryx_u64(0x0, x284, x286, &x295);
  { uint64_t x298; uint8_t x299 = _addcarryx_u64(x296, x287, x289, &x298);
  { uint64_t x301; uint8_t x302 = _addcarryx_u64(x299, x290, x292, &x301);
  { uint8_t x303 = (x302 + x293);
  { uint64_t _; uint8_t x306 = _addcarryx_u64(0x0, x265, x283, &_);
  { uint64_t x308; uint8_t x309 = _addcarryx_u64(x306, x268, x295, &x308);
  { uint64_t x311; uint8_t x312 = _addcarryx_u64(x309, x271, x298, &x311);
  { uint64_t x314; uint8_t x315 = _addcarryx_u64(x312, x274, x301, &x314);
  { uint64_t x317; uint8_t x318 = _addcarryx_u64(x315, x277, x303, &x317);
  { uint8_t x319 = (x318 + x278);
  { uint64_t x321; uint8_t x322 = _subborrow_u64(0x0, x308, 0xffffffffffffffdfL, &x321);
  { uint64_t x324; uint8_t x325 = _subborrow_u64(x322, x311, 0xffffffffffffffffL, &x324);
  { uint64_t x327; uint8_t x328 = _subborrow_u64(x325, x314, 0xffffffffffffffffL, &x327);
  { uint64_t x330; uint8_t x331 = _subborrow_u64(x328, x317, 0x3, &x330);
  { uint64_t _; uint8_t x334 = _subborrow_u64(x331, x319, 0x0, &_);
  { uint64_t x335 = cmovznz64(x334, x330, x317);
  { uint64_t x336 = cmovznz64(x334, x327, x314);
  { uint64_t x337 = cmovznz64(x334, x324, x311);
  { uint64_t x338 = cmovznz64(x334, x321, x308);
  out[0] = x338;
  out[1] = x337;
  out[2] = x336;
  out[3] = x335;
  }}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}}
}