jmp to self instruction compiled by gcc4.4.6-3

Question

I have a C++ project, run rightly compiled by gcc-4.1.2-46 and gcc-4.4.5-6

But it has an abnormal dead loop while compiled by gcc-4.4.6-3 using -O2.

I use gdb to

Answer 1

I suspect the problem is in your loop in BitLock::lock:

    while (true)
    {
      uint8_t ov = bits_[byte_index];
      if (ov & BIT_MASKS[bit_index])
      {
        continue;
      }
      if (ov == ATOMIC_CAS(&(bits_[byte_index]), ov, ov | BIT_MASKS[bit_index]))
      {
        break;
      }
    }

The __sync_val_compare_and_swap in your ATOMIC_CAS macro will act as memory barrier, preventing the compiler/processor from speculating loads across it, but it won't do anything about loads before it, so the condition ov & BIT_MASKS[bit_index] could be optimized based on the value of bits_[byte_index] before the __sync_val_compare_and_swap call and thus resulting in an infinite loop.

Try the following, instead:

    while (true)
    {
      uint8_t ov = bits_[byte_index];
      if (ov & BIT_MASKS[bit_index])
      {
        __sync_synchronize();  // or define ATOMIC_SYNC if you prefer
        continue;
      }
      if (ov == ATOMIC_CAS(&(bits_[byte_index]), ov, ov | BIT_MASKS[bit_index]))
      {
        break;
      }
    }

The __sync_synchronize memory barrier will prevent the loop from degenerating into a trivial one.