I am currently working on a project, where every cycle counts. While profiling my application I discovered that the overhead of some inner loop is quite high, because they consist of just a few machine instruction. Additionally the number of iterations in these loops is known at compile time.
So I thought instead of manually unrolling the loop with copy & paste I could use macros to unroll the loop at compile time so that it can be easily modified later.
What I image is something like this:
#define LOOP_N_TIMES(N, CODE) <insert magic here>
So that I can replace for (int i = 0; i < N, ++i) { do_stuff(); }
with:
#define INNER_LOOP_COUNT 4
LOOP_N_TIMES(INNER_LOOP_COUNT, do_stuff();)
And it unrolls itself to:
do_stuff(); do_stuff(); do_stuff(); do_stuff();
Since the C preprocessor is still a mystery to me most of the time, I have no idea how to accomplish this, but I know it must be possible because Boost seems to have a BOOST_PP_REPEAT
macros. Unfortunately I can't use Boost for this project.
You can use templates to unroll. See the disassembly for the sample Live on Godbolt
But -funroll-loops
has the same effect for this sample.
template <unsigned N> struct faux_unroll {
template <typename F> static void call(F const& f) {
f();
faux_unroll<N-1>::call(f);
}
};
template <> struct faux_unroll<0u> {
template <typename F> static void call(F const&) {}
};
#include <iostream>
#include <cstdlib>
int main() {
srand(time(0));
double r = 0;
faux_unroll<10>::call([&] { r += 1.0/rand(); });
std::cout << r;
}