Depends on compiler options. x86 floating point is 80-bit internally, but 64-bit in memory. Compile the same function with different optimization, for a different CPU, or whatever and get different answers. For gcc, the answer is to use -ffloat-store. But you have to remember to use it.