long print(const void *buf, long count) {
#if defined(__x86_64__)
    asm volatile (
        "syscall"
        :
        : "a"(1), "D"(1), "S"(buf), "d"(count)  // RDI, RSI, RDX are used in x86-64 for write args
        : "rcx", "r11", "memory"
    );
    #elif defined(__aarch64__)
    asm ( "mov x0, #1\n"        // Move the exit status into register x0
          "mov x1, %0\n"       // Syscall number for 'exit' is 93 in AArch64
          "mov x2, %1\n"       // Syscall number for 'exit' is 93 in AArch64
          "mov x8, #64\n"       // Syscall number for 'exit' is 93 in AArch64
          "svc #0\n"            // Make the syscall
          :                      // No output operands
          : "r" (buf), "r"(count)         // Input operand: status
          : "x0", "x8"           // Clobbered registers
    );
    #elif defined(__mips__)
  asm ( "li $a0, 1\n"
	"move $a1, %0\n"
	"move $a2, %1\n"
	"li $v0, 4004\n"
	"syscall"
	:
	: "r"(buf), "r"(count)
	: "a0", "a1", "a2", "v0" //TODO: temp registers clobbered by syscall should probably also be listed but this fn returns right away so it should be fine?
      );
    #endif
    return 0;
}

#define CLONE_CHILD_SETTID   0x1000000
#define CLONE_CHILD_CLEARTID 0x200000
#if defined(__aarch64__)
long int fork() {
  long int rtn;
  long int flags = CLONE_CHILD_CLEARTID | CLONE_CHILD_SETTID;
    asm volatile (
      // Assembly Instructions
      "mov x8, #220\n"  // Syscall number for 'clone' is 220 in AArch64
      "mov x0, %1\n"    // clone flags
      "mov x1, #0\n"    // child stack pointer
      "mov x2, #0\n"    // parent_tidptr
      "mov x3, #0\n"    // tls
      "mov x4, #0\n"    // child_tidptr
      "svc #0\n"        // Make the syscall
      "mov %0, x0\n"    // save return value
      // Output operands
      : "=r" (rtn)
      // Input operand
      : "r" (flags)
      // Clobbered registers
      : "x0", "x1", "x2", "x3", "x4", "x8"
    );
  return rtn;
}
#elif defined(__x86_64__)
long int fork() {
  long int rtn;
  /*
  long int flags = CLONE_CHILD_CLEARTID | CLONE_CHILD_SETTID;
  asm volatile(
    "movq %%rax, 56\n"   // syscall number for 'clone' is 56 for x86-64
    "movq %%rdi, %1\n"    // clone flags
    "movq %%rsi, 0\n"    // child stack pointer
    "movq %%rdx, 0\n"    // parent tidptr
    "movq %%r10, 0\n"    // child tidptr
    "movq %%r8,  0\n"     // tls
    "syscall\n"         // syscall
    "movq %0, %%rax\n"    // save return value
    : "=r"(rtn)
    : "r"(flags)
    : "rax", "rdi", "rsi", "rdx", "r10", "r8", "r9", "r10", "r11"
  );
  */
  asm volatile(
    "syscall\n"         // syscall
    "movq %0, %%rax\n"    // save return value
    : "=r"(rtn)
    : "a"(57)
    :
  );
  return rtn;
}
#endif

void exit(int status) {
#if defined(__x86_64__)
  asm volatile(
    "syscall"
    :
    : "a"(60), "D"(status)
    :
    );
    #elif defined(__aarch64__)
  asm ( "mov x0, %0\n"        // Move the exit status into register x0
        "mov x8, #93\n"       // Syscall number for 'exit' is 93 in AArch64
        "svc #0\n"            // Make the syscall
        :                      // No output operands
        : "r" ((long)status)         // Input operand: status
        : "x0", "x8"           // Clobbered registers
  );
    #elif defined(__mips__)
  asm ( "move $a0, %0\n"
	"li $v0, 4001\n"
	"syscall"
	:
	: "r" (status)
	: "a0", "v0"
      );
    #endif
  for(;;);
}


int main() {
#if defined(__x86_64__) || defined(__aarch64__)
  long int pid = fork();
  char msg[17] = {' '};
  msg[16] = '\n';
  for(int i = 0; i < 16; ++i) {
    char nibble = (pid >> ((15 - i) * 4)) & 0xf;
    char c;
    if (nibble > 9) {c = nibble + '7';}
    else {c = nibble + '0';}
    msg[i] = c;
  }
  print(msg, 17);
  #else
  char *msg = "Hello World!\n";
  print(msg, 13);
#endif
  return 69;
}
void __libc_start_main() {exit(main());}

void _start() {
  __libc_start_main();
}

void __start() {
  _start();
}
/*
#if defined(__x86_64__)
        printf("Compiled for x86_64 architecture\n");
    #elif defined(__i386__)
        printf("Compiled for x86 architecture\n");
    #elif defined(__arm__)
        printf("Compiled for ARM architecture\n");
    #elif defined(__aarch64__)
        printf("Compiled for ARM64 architecture\n");
    #elif defined(__mips__)
        printf("Compiled for ARM architecture\n");
    #else
        printf("Unknown architecture\n");
    #endif
*/