Possibility to force compiler roll out a for loop

Question

In order to increase performance at runtime is there a possibility to force the compiler to role out such a for loop with compiler known count of iterations?
Such that this code.

template <int dim>
void foo()
{
   ...
   int temp[i];
   for(int i=0; i<dim; i++)
   {
      temp[i] = ...;
      ...
   }
   ...
}

is build as it would be

template <int dim>
void foo()
{
   ...
   int temp[i];

      temp[0] = ...;
      temp[1] = ...;
      temp[2] = ...;
      (...)
      temp[dim-1] = ...;
      ...
   }
   ...
}

You don't need to do it. It will do it on it's own since is a simple case. — VAndrei, Nov 16 '14 at 13:36
Check that one out for an old school way of doing loop unrolling: http://stackoverflow.com/questions/514118/how-does-duffs-device-work — didierc, Nov 16 '14 at 13:42

score 5 · Answer 1 · edited May 23 '17 at 12:20

Although I don't see why you wouldn't let the compiler decice whether unrolling is beneficial to the performance (which it often isn't), you can, to be compiler-independent, use variadic templates with pack-expansions. E.g. in C++14:

namespace detail
{
    template <std::size_t... i>
    void foo( std::index_sequence<i...> )
    {
        int temp[sizeof...(i)];
        (void)std::initializer_list<int>{ (temp[i] = i)... };
    }
}

template <std::size_t dim>
void foo()
{
    detail::foo( std::make_index_sequence<dim>() );
}

Demo. Otherwise you could use Pragmas. This answer describes how to do that for GCC.

Possibility to force compiler roll out a for loop

1 Answers1