DATA ·andMask<>+0x00(SB)/8, $0x00000000000000ff
DATA ·andMask<>+0x08(SB)/8, $0x0000000000000000
DATA ·andMask<>+0x10(SB)/8, $0x000000000000ffff
DATA ·andMask<>+0x18(SB)/8, $0x0000000000000000
DATA ·andMask<>+0x20(SB)/8, $0x0000000000ffffff
DATA ·andMask<>+0x28(SB)/8, $0x0000000000000000
DATA ·andMask<>+0x30(SB)/8, $0x00000000ffffffff
DATA ·andMask<>+0x38(SB)/8, $0x0000000000000000
DATA ·andMask<>+0x40(SB)/8, $0x000000ffffffffff

GLOBL ·IEEEConst(SB),RODATA,$4336

	/* Barrett constant m - (4^32)/n */
DATA ·IEEEBarConst(SB)/8,$0x00000001f7011641
DATA ·IEEEBarConst+8(SB)/8,$0x0000000000000000
DATA ·IEEEBarConst+16(SB)/8,$0x00000001db710641
DATA ·IEEEBarConst+24(SB)/8,$0x0000000000000000
GLOBL ·IEEEBarConst(SB),RODATA,$32

	/* Reduce 262144 kbits to 1024 bits */
	/* x^261184 mod p(x), x^261120 mod p(x) */

    }
    buffer_data_map_.erase(it);
    it = buffer_data_map_.begin();
    buffer_data_exported_ = true;
  }
  // pad 16 bytes for the last buffer for XNNPack
  result.Append("\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0");
  // pad to be 16 bytes aligned
  {
    std::string pad(kFbAlignment - result.size() % kFbAlignment, '\0');
    result.Append(std::move(pad));
  }

  for (auto& it : custom_op_data_map_) {