#include <benchmark/benchmark.h>
#include <numeric>
#include <random>

int suma1(int *begin, int *end)
{
    int total = 0;
    asm volatile("0: cmp %0,%2     \n"
                 "   je 1f         \n"
                 "   add $4,%0     \n" // esto se
                 "   add -4(%0),%1 \n" // ve raro
                 "   jmp 0b        \n"
                 "1:               \n"
                 : "+&r"(begin), "+&r"(total)
                 : "r"(end)
                 : "cc", "memory");
    return total;
}

int suma2(int *begin, int *end)
{
    int total = 0;
    asm volatile("0: cmp %0,%2     \n"
                 "   je 1f         \n"
                 "   add (%0),%1   \n" // ahora no,
                 "   add $4,%0     \n" // verdad?
                 "   jmp 0b        \n"
                 "1:               \n"
                 : "+&r"(begin), "+&r"(total)
                 : "r"(end)
                 : "cc", "memory");
    return total;
}
template<auto f> static void test(benchmark::State &state)
{
    const std::size_t N = 1'000'000;
    int v[N];
    long int suma = 0;

    std::iota(v, v + N, 0);

    for (auto _ : state)
        benchmark::DoNotOptimize(suma += f(v, v + N));

    state.SetLabel(std::to_string(suma));
}

BENCHMARK(test<suma1>)->Iterations(2'000);
BENCHMARK(test<suma2>)->Iterations(2'000);

BENCHMARK_MAIN();