On 12 April 2010 22:57, Chad Nelson <chad.thecomfychair@gmail.com> wrote:
Right now, the XInt library uses a pointer to a (privately defined) struct. The struct has several items, including a small array of "quick digits" and a pointer to a dynamically allocated array of digit_t types, both for containing the magnitude of the integer.
The most important optimization for me is the SBO. If the magnitude of the int in question is less than 2**30, it should take no dynamic storage at all. (In addition to space gains, this is a big win for exception-safety too.)
It's an old C trick, apparently legitimized in the C99 standard as the "struct hack". For those not familiar with it, the idea is that I'd allocate a single array of digit_t types, of the required size plus sizeof(new_data_t), then just typecast it as a pointer of type new_data_t.
Does anyone know if there's any reason not to do this? Maybe some portability or alignment problem with it?
I believe the official C way is that the last element is an array without a listed length (digits_type digits_[];), so you malloc once for the whole type and use the array being sure not to overrun the space you allocated. I'm not convinced such tricks are officially allowed in C++, given that the new/delete model doesn't offer any way to use them. Since you just have integral types, why not just allocate new digits_type[allocated+2] and know that index 0 is used and index 1 is allocated? I doubt anyone would be too concerned about not being able to store 2**37-bit integers.