#ifdef __GNUC__
#pragma once
#include <inttypes.h>

static inline uint64_t add1_inline (uint64_t* arg0, uint64_t* arg1, uint64_t arg2) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t arg2_r asm("rdx") = arg2;
  register uint64_t carry_r asm("rax");

  __asm__ __volatile__(
    "  xor %%r8, %%r8;"
    "  xor %%r9, %%r9;"
    "  xor %%r10, %%r10;"
    "  xor %%r11, %%r11;"
    "  xor %%rax, %%rax;"
    "  addq 0(%%rsi), %%rdx;"
    "  movq %%rdx, 0(%%rdi);"
    "  adcxq 8(%%rsi), %%r8;"
    "  movq %%r8, 8(%%rdi);"
    "  adcxq 16(%%rsi), %%r9;"
    "  movq %%r9, 16(%%rdi);"
    "  adcxq 24(%%rsi), %%r10;"
    "  movq %%r10, 24(%%rdi);"
    "  adcx %%r11, %%rax;"
  : "+r" (arg2_r), "=r" (carry_r)
  : "r" (arg0_r), "r" (arg1_r)
  : "%r8", "%r9", "%r10", "%r11", "memory", "cc"
  );

return carry_r;
}

static inline void fadd_inline (uint64_t* arg0, uint64_t* arg1, uint64_t* arg2) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;

  __asm__ __volatile__(
    "  movq 0(%%rdx), %%r8;"
    "  addq 0(%%rsi), %%r8;"
    "  movq 8(%%rdx), %%r9;"
    "  adcxq 8(%%rsi), %%r9;"
    "  movq 16(%%rdx), %%r10;"
    "  adcxq 16(%%rsi), %%r10;"
    "  movq 24(%%rdx), %%r11;"
    "  adcxq 24(%%rsi), %%r11;"
    "  mov $0, %%rax;"
    "  mov $38, %%rdx;"
    "  cmovc %%rdx, %%rax;"
    "  xor %%rcx, %%rcx;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 8(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 24(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
  : "+r" (arg2_r)
  : "r" (arg0_r), "r" (arg1_r)
  : "%rax", "%rcx", "%r8", "%r9", "%r10", "%r11", "memory", "cc"
  );
}

static inline void fsub_inline (uint64_t* arg0, uint64_t* arg1, uint64_t* arg2) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;

  __asm__ __volatile__(
    "  movq 0(%%rsi), %%r8;"
    "  subq 0(%%rdx), %%r8;"
    "  movq 8(%%rsi), %%r9;"
    "  sbbq 8(%%rdx), %%r9;"
    "  movq 16(%%rsi), %%r10;"
    "  sbbq 16(%%rdx), %%r10;"
    "  movq 24(%%rsi), %%r11;"
    "  sbbq 24(%%rdx), %%r11;"
    "  mov $0, %%rax;"
    "  mov $38, %%rcx;"
    "  cmovc %%rcx, %%rax;"
    "  sub %%rax, %%r8;"
    "  sbb $0, %%r9;"
    "  sbb $0, %%r10;"
    "  sbb $0, %%r11;"
    "  mov $0, %%rax;"
    "  cmovc %%rcx, %%rax;"
    "  sub %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
    "  movq %%r9, 8(%%rdi);"
    "  movq %%r10, 16(%%rdi);"
    "  movq %%r11, 24(%%rdi);"
  : 
  : "r" (arg0_r), "r" (arg1_r), "r" (arg2_r)
  : "%rax", "%rcx", "%r8", "%r9", "%r10", "%r11", "memory", "cc"
  );
}

static inline void fmul_inline (uint64_t* arg0, uint64_t* arg1, uint64_t* arg2, uint64_t* arg3) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;
  register uint64_t* arg3_r asm("rcx") = arg3;

  __asm__ __volatile__(
    "  mov %%rdx, %%r15;"
    "  movq 0(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  movq %%r8, 0(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  movq %%r10, 8(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  movq 8(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 8(%%rdi), %%r8;"
    "  movq %%r8, 8(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  mov $0, %%r8;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq 16(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 16(%%rdi), %%r8;"
    "  movq %%r8, 16(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 24(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  mov $0, %%r8;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq 24(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 24(%%rdi), %%r8;"
    "  movq %%r8, 24(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 32(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  movq %%r12, 40(%%rdi);"
    "  mov $0, %%r8;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  movq %%r14, 48(%%rdi);"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq %%rax, 56(%%rdi);"
    "  mov %%rdi, %%rsi;"
    "  mov %%r15, %%rdi;"
    "  mov $38, %%rdx;"
    "  mulxq 32(%%rsi), %%r8, %%r13;"
    "  xor %%rcx, %%rcx;"
    "  adoxq 0(%%rsi), %%r8;"
    "  mulxq 40(%%rsi), %%r9, %%r12;"
    "  adcx %%r13, %%r9;"
    "  adoxq 8(%%rsi), %%r9;"
    "  mulxq 48(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  adoxq 16(%%rsi), %%r10;"
    "  mulxq 56(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adoxq 24(%%rsi), %%r11;"
    "  adcx %%rcx, %%rax;"
    "  adox %%rcx, %%rax;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 8(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 24(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
  : "+r" (arg0_r), "+r" (arg1_r), "+r" (arg2_r), "+r" (arg3_r)
  : 
  : "%rax", "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15", "memory", "cc"
  );
}

static inline void fmul2_inline (uint64_t* arg0, uint64_t* arg1, uint64_t* arg2, uint64_t* arg3) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;
  register uint64_t* arg3_r asm("rcx") = arg3;

  __asm__ __volatile__(
    "  mov %%rdx, %%r15;"
    "  movq 0(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  movq %%r8, 0(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  movq %%r10, 8(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  movq 8(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 8(%%rdi), %%r8;"
    "  movq %%r8, 8(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  mov $0, %%r8;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq 16(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 16(%%rdi), %%r8;"
    "  movq %%r8, 16(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 24(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  mov $0, %%r8;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq 24(%%rsi), %%rdx;"
    "  mulxq 0(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 24(%%rdi), %%r8;"
    "  movq %%r8, 24(%%rdi);"
    "  mulxq 8(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 32(%%rdi);"
    "  mulxq 16(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  movq %%r12, 40(%%rdi);"
    "  mov $0, %%r8;"
    "  mulxq 24(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  movq %%r14, 48(%%rdi);"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq %%rax, 56(%%rdi);"
    "  movq 32(%%rsi), %%rdx;"
    "  mulxq 32(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  movq %%r8, 64(%%rdi);"
    "  mulxq 40(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  movq %%r10, 72(%%rdi);"
    "  mulxq 48(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  mulxq 56(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  movq 40(%%rsi), %%rdx;"
    "  mulxq 32(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 72(%%rdi), %%r8;"
    "  movq %%r8, 72(%%rdi);"
    "  mulxq 40(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 80(%%rdi);"
    "  mulxq 48(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  mov $0, %%r8;"
    "  mulxq 56(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq 48(%%rsi), %%rdx;"
    "  mulxq 32(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 80(%%rdi), %%r8;"
    "  movq %%r8, 80(%%rdi);"
    "  mulxq 40(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 88(%%rdi);"
    "  mulxq 48(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  mov $0, %%r8;"
    "  mulxq 56(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq 56(%%rsi), %%rdx;"
    "  mulxq 32(%%rcx), %%r8, %%r9;"
    "  xor %%r10, %%r10;"
    "  adcxq 88(%%rdi), %%r8;"
    "  movq %%r8, 88(%%rdi);"
    "  mulxq 40(%%rcx), %%r10, %%r11;"
    "  adox %%r9, %%r10;"
    "  adcx %%r12, %%r10;"
    "  movq %%r10, 96(%%rdi);"
    "  mulxq 48(%%rcx), %%r12, %%r13;"
    "  adox %%r11, %%r12;"
    "  adcx %%r14, %%r12;"
    "  movq %%r12, 104(%%rdi);"
    "  mov $0, %%r8;"
    "  mulxq 56(%%rcx), %%r14, %%rdx;"
    "  adox %%r13, %%r14;"
    "  adcx %%rax, %%r14;"
    "  movq %%r14, 112(%%rdi);"
    "  mov $0, %%rax;"
    "  adox %%rdx, %%rax;"
    "  adcx %%r8, %%rax;"
    "  movq %%rax, 120(%%rdi);"
    "  mov %%rdi, %%rsi;"
    "  mov %%r15, %%rdi;"
    "  mov $38, %%rdx;"
    "  mulxq 32(%%rsi), %%r8, %%r13;"
    "  xor %%rcx, %%rcx;"
    "  adoxq 0(%%rsi), %%r8;"
    "  mulxq 40(%%rsi), %%r9, %%r12;"
    "  adcx %%r13, %%r9;"
    "  adoxq 8(%%rsi), %%r9;"
    "  mulxq 48(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  adoxq 16(%%rsi), %%r10;"
    "  mulxq 56(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adoxq 24(%%rsi), %%r11;"
    "  adcx %%rcx, %%rax;"
    "  adox %%rcx, %%rax;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 8(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 24(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
    "  mov $38, %%rdx;"
    "  mulxq 96(%%rsi), %%r8, %%r13;"
    "  xor %%rcx, %%rcx;"
    "  adoxq 64(%%rsi), %%r8;"
    "  mulxq 104(%%rsi), %%r9, %%r12;"
    "  adcx %%r13, %%r9;"
    "  adoxq 72(%%rsi), %%r9;"
    "  mulxq 112(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  adoxq 80(%%rsi), %%r10;"
    "  mulxq 120(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adoxq 88(%%rsi), %%r11;"
    "  adcx %%rcx, %%rax;"
    "  adox %%rcx, %%rax;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 40(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 48(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 56(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 32(%%rdi);"
  : "+r" (arg0_r), "+r" (arg1_r), "+r" (arg2_r), "+r" (arg3_r)
  : 
  : "%rax", "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15", "memory", "cc"
  );
}

static inline void fmul1_inline (uint64_t* arg0, uint64_t* arg1, uint64_t arg2) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t arg2_r asm("rdx") = arg2;

  __asm__ __volatile__(
    "  mulxq 0(%%rsi), %%r8, %%rcx;"
    "  mulxq 8(%%rsi), %%r9, %%r12;"
    "  add %%rcx, %%r9;"
    "  mov $0, %%rcx;"
    "  mulxq 16(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  mulxq 24(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adcx %%rcx, %%rax;"
    "  mov $38, %%rdx;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 8(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 24(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
  : "+r" (arg2_r)
  : "r" (arg0_r), "r" (arg1_r)
  : "%rax", "%rcx", "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "memory", "cc"
  );
}

static inline void cswap2_inline (uint64_t arg0, uint64_t* arg1, uint64_t* arg2) {
  register uint64_t arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;

  __asm__ __volatile__(
    "  add $18446744073709551615, %%rdi;"
    "  movq 0(%%rsi), %%r8;"
    "  movq 0(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 0(%%rsi);"
    "  movq %%r9, 0(%%rdx);"
    "  movq 8(%%rsi), %%r8;"
    "  movq 8(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 8(%%rsi);"
    "  movq %%r9, 8(%%rdx);"
    "  movq 16(%%rsi), %%r8;"
    "  movq 16(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 16(%%rsi);"
    "  movq %%r9, 16(%%rdx);"
    "  movq 24(%%rsi), %%r8;"
    "  movq 24(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 24(%%rsi);"
    "  movq %%r9, 24(%%rdx);"
    "  movq 32(%%rsi), %%r8;"
    "  movq 32(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 32(%%rsi);"
    "  movq %%r9, 32(%%rdx);"
    "  movq 40(%%rsi), %%r8;"
    "  movq 40(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 40(%%rsi);"
    "  movq %%r9, 40(%%rdx);"
    "  movq 48(%%rsi), %%r8;"
    "  movq 48(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 48(%%rsi);"
    "  movq %%r9, 48(%%rdx);"
    "  movq 56(%%rsi), %%r8;"
    "  movq 56(%%rdx), %%r9;"
    "  mov %%r8, %%r10;"
    "  cmovc %%r9, %%r8;"
    "  cmovc %%r10, %%r9;"
    "  movq %%r8, 56(%%rsi);"
    "  movq %%r9, 56(%%rdx);"
  : "+r" (arg0_r)
  : "r" (arg1_r), "r" (arg2_r)
  : "%r8", "%r9", "%r10", "memory", "cc"
  );
}

static inline void fsqr_inline (uint64_t* arg0, uint64_t* arg1, uint64_t* arg2) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;

  __asm__ __volatile__(
    "  mov %%rdx, %%rbx;"
    "  movq 0(%%rsi), %%rdx;"
    "  mulxq 8(%%rsi), %%r8, %%r14;"
    "  xor %%r15, %%r15;"
    "  mulxq 16(%%rsi), %%r9, %%r10;"
    "  adcx %%r14, %%r9;"
    "  mulxq 24(%%rsi), %%rax, %%rcx;"
    "  adcx %%rax, %%r10;"
    "  movq 24(%%rsi), %%rdx;"
    "  mulxq 8(%%rsi), %%r11, %%r12;"
    "  adcx %%rcx, %%r11;"
    "  mulxq 16(%%rsi), %%rax, %%r13;"
    "  adcx %%rax, %%r12;"
    "  movq 8(%%rsi), %%rdx;"
    "  adcx %%r15, %%r13;"
    "  mulxq 16(%%rsi), %%rax, %%rcx;"
    "  mov $0, %%r14;"
    "  xor %%r15, %%r15;"
    "  adox %%rax, %%r10;"
    "  adcx %%r8, %%r8;"
    "  adox %%rcx, %%r11;"
    "  adcx %%r9, %%r9;"
    "  adox %%r15, %%r12;"
    "  adcx %%r10, %%r10;"
    "  adox %%r15, %%r13;"
    "  adcx %%r11, %%r11;"
    "  adox %%r15, %%r14;"
    "  adcx %%r12, %%r12;"
    "  adcx %%r13, %%r13;"
    "  adcx %%r14, %%r14;"
    "  movq 0(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  movq %%rax, 0(%%rdi);"
    "  add %%rcx, %%r8;"
    "  movq %%r8, 8(%%rdi);"
    "  movq 8(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r9;"
    "  movq %%r9, 16(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 24(%%rdi);"
    "  movq 16(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r11;"
    "  movq %%r11, 32(%%rdi);"
    "  adcx %%rcx, %%r12;"
    "  movq %%r12, 40(%%rdi);"
    "  movq 24(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r13;"
    "  movq %%r13, 48(%%rdi);"
    "  adcx %%rcx, %%r14;"
    "  movq %%r14, 56(%%rdi);"
    "  mov %%rdi, %%rsi;"
    "  mov %%rbx, %%rdi;"
    "  mov $38, %%rdx;"
    "  mulxq 32(%%rsi), %%r8, %%r13;"
    "  xor %%rcx, %%rcx;"
    "  adoxq 0(%%rsi), %%r8;"
    "  mulxq 40(%%rsi), %%r9, %%r12;"
    "  adcx %%r13, %%r9;"
    "  adoxq 8(%%rsi), %%r9;"
    "  mulxq 48(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  adoxq 16(%%rsi), %%r10;"
    "  mulxq 56(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adoxq 24(%%rsi), %%r11;"
    "  adcx %%rcx, %%rax;"
    "  adox %%rcx, %%rax;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 8(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 24(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
  : "+r" (arg0_r), "+r" (arg1_r), "+r" (arg2_r)
  : 
  : "%rax", "%rbx", "%rcx", "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15", "memory", "cc"
  );
}

static inline void fsqr2_inline (uint64_t* arg0, uint64_t* arg1, uint64_t* arg2) {
  register uint64_t* arg0_r asm("rdi") = arg0;
  register uint64_t* arg1_r asm("rsi") = arg1;
  register uint64_t* arg2_r asm("rdx") = arg2;

  __asm__ __volatile__(
    "  mov %%rdx, %%rbx;"
    "  movq 0(%%rsi), %%rdx;"
    "  mulxq 8(%%rsi), %%r8, %%r14;"
    "  xor %%r15, %%r15;"
    "  mulxq 16(%%rsi), %%r9, %%r10;"
    "  adcx %%r14, %%r9;"
    "  mulxq 24(%%rsi), %%rax, %%rcx;"
    "  adcx %%rax, %%r10;"
    "  movq 24(%%rsi), %%rdx;"
    "  mulxq 8(%%rsi), %%r11, %%r12;"
    "  adcx %%rcx, %%r11;"
    "  mulxq 16(%%rsi), %%rax, %%r13;"
    "  adcx %%rax, %%r12;"
    "  movq 8(%%rsi), %%rdx;"
    "  adcx %%r15, %%r13;"
    "  mulxq 16(%%rsi), %%rax, %%rcx;"
    "  mov $0, %%r14;"
    "  xor %%r15, %%r15;"
    "  adox %%rax, %%r10;"
    "  adcx %%r8, %%r8;"
    "  adox %%rcx, %%r11;"
    "  adcx %%r9, %%r9;"
    "  adox %%r15, %%r12;"
    "  adcx %%r10, %%r10;"
    "  adox %%r15, %%r13;"
    "  adcx %%r11, %%r11;"
    "  adox %%r15, %%r14;"
    "  adcx %%r12, %%r12;"
    "  adcx %%r13, %%r13;"
    "  adcx %%r14, %%r14;"
    "  movq 0(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  movq %%rax, 0(%%rdi);"
    "  add %%rcx, %%r8;"
    "  movq %%r8, 8(%%rdi);"
    "  movq 8(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r9;"
    "  movq %%r9, 16(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 24(%%rdi);"
    "  movq 16(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r11;"
    "  movq %%r11, 32(%%rdi);"
    "  adcx %%rcx, %%r12;"
    "  movq %%r12, 40(%%rdi);"
    "  movq 24(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r13;"
    "  movq %%r13, 48(%%rdi);"
    "  adcx %%rcx, %%r14;"
    "  movq %%r14, 56(%%rdi);"
    "  movq 32(%%rsi), %%rdx;"
    "  mulxq 40(%%rsi), %%r8, %%r14;"
    "  xor %%r15, %%r15;"
    "  mulxq 48(%%rsi), %%r9, %%r10;"
    "  adcx %%r14, %%r9;"
    "  mulxq 56(%%rsi), %%rax, %%rcx;"
    "  adcx %%rax, %%r10;"
    "  movq 56(%%rsi), %%rdx;"
    "  mulxq 40(%%rsi), %%r11, %%r12;"
    "  adcx %%rcx, %%r11;"
    "  mulxq 48(%%rsi), %%rax, %%r13;"
    "  adcx %%rax, %%r12;"
    "  movq 40(%%rsi), %%rdx;"
    "  adcx %%r15, %%r13;"
    "  mulxq 48(%%rsi), %%rax, %%rcx;"
    "  mov $0, %%r14;"
    "  xor %%r15, %%r15;"
    "  adox %%rax, %%r10;"
    "  adcx %%r8, %%r8;"
    "  adox %%rcx, %%r11;"
    "  adcx %%r9, %%r9;"
    "  adox %%r15, %%r12;"
    "  adcx %%r10, %%r10;"
    "  adox %%r15, %%r13;"
    "  adcx %%r11, %%r11;"
    "  adox %%r15, %%r14;"
    "  adcx %%r12, %%r12;"
    "  adcx %%r13, %%r13;"
    "  adcx %%r14, %%r14;"
    "  movq 32(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  movq %%rax, 64(%%rdi);"
    "  add %%rcx, %%r8;"
    "  movq %%r8, 72(%%rdi);"
    "  movq 40(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r9;"
    "  movq %%r9, 80(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 88(%%rdi);"
    "  movq 48(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r11;"
    "  movq %%r11, 96(%%rdi);"
    "  adcx %%rcx, %%r12;"
    "  movq %%r12, 104(%%rdi);"
    "  movq 56(%%rsi), %%rdx;"
    "  mulx %%rdx, %%rax, %%rcx;"
    "  adcx %%rax, %%r13;"
    "  movq %%r13, 112(%%rdi);"
    "  adcx %%rcx, %%r14;"
    "  movq %%r14, 120(%%rdi);"
    "  mov %%rdi, %%rsi;"
    "  mov %%rbx, %%rdi;"
    "  mov $38, %%rdx;"
    "  mulxq 32(%%rsi), %%r8, %%r13;"
    "  xor %%rcx, %%rcx;"
    "  adoxq 0(%%rsi), %%r8;"
    "  mulxq 40(%%rsi), %%r9, %%r12;"
    "  adcx %%r13, %%r9;"
    "  adoxq 8(%%rsi), %%r9;"
    "  mulxq 48(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  adoxq 16(%%rsi), %%r10;"
    "  mulxq 56(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adoxq 24(%%rsi), %%r11;"
    "  adcx %%rcx, %%rax;"
    "  adox %%rcx, %%rax;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 8(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 16(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 24(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 0(%%rdi);"
    "  mov $38, %%rdx;"
    "  mulxq 96(%%rsi), %%r8, %%r13;"
    "  xor %%rcx, %%rcx;"
    "  adoxq 64(%%rsi), %%r8;"
    "  mulxq 104(%%rsi), %%r9, %%r12;"
    "  adcx %%r13, %%r9;"
    "  adoxq 72(%%rsi), %%r9;"
    "  mulxq 112(%%rsi), %%r10, %%r13;"
    "  adcx %%r12, %%r10;"
    "  adoxq 80(%%rsi), %%r10;"
    "  mulxq 120(%%rsi), %%r11, %%rax;"
    "  adcx %%r13, %%r11;"
    "  adoxq 88(%%rsi), %%r11;"
    "  adcx %%rcx, %%rax;"
    "  adox %%rcx, %%rax;"
    "  imul %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  adcx %%rcx, %%r9;"
    "  movq %%r9, 40(%%rdi);"
    "  adcx %%rcx, %%r10;"
    "  movq %%r10, 48(%%rdi);"
    "  adcx %%rcx, %%r11;"
    "  movq %%r11, 56(%%rdi);"
    "  mov $0, %%rax;"
    "  cmovc %%rdx, %%rax;"
    "  add %%rax, %%r8;"
    "  movq %%r8, 32(%%rdi);"
  : "+r" (arg0_r), "+r" (arg1_r), "+r" (arg2_r)
  : 
  : "%rax", "%rbx", "%rcx", "%r8", "%r9", "%r10", "%r11", "%r12", "%r13", "%r14", "%r15", "memory", "cc"
  );
}

#endif