Add COW element access on transients?

[arximboldi]

This would allow to do std::sort directly saving some copies. Not sure whether it is a good idea though. There are some drawbacks.

[kevin--]

attempting to do std::sort on a immer::vector -- is this just out of the question right now, or must you copy your contents to an STL container to perform std::sort?

[arximboldi]

Yes, right now one has to copy the elements into a std::vector. I've considered adding COW element access, but I actually doubt it will be faster in this case. Since every element is touched, more than once actually, I think the relatively high constant on random access on immer::vector is higher than copying it all into a std::vector and back (which when done properly should be pretty fast!).

[arximboldi]

Which I realize, the fastest way to convert an immer::vector to std::vector is maybe not obvious. Actually, I'm not sure which of these two versions is actually faster:

template <typename T>
std::vector<T> to_std(const immer::vector<T>& v)
{
    auto r = std::vector<T>{};
    r.reserve(v.size());
    immer::copy(v, std::back_inserter(v));
    return r;
}

std::vector<int> to_std(const immer::vector<int>& v)
{
    auto r = std::vector<int>{};
    r.reserve(v.size());
    immer::for_each_chunk(v, [&](auto b, auto e) {
        r.insert(r.end(), b, e);
    });
    return r;
}

I'm sure now: https://godbolt.org/z/d5n1Wj8MK

[arximboldi]

This makes me think, maybe this is such a common thing to do that a family of to_std and from_std functions would make sense in Immer!

[kevin--]

yes that makes a lot of sense to provide the std conversion routines. I have written the from_std equivalent for interop with a serialization library... For now I am maintaining a sorted collection in flex_vector as described in #105 using std::lower_bound

Thanks also for the tip on BENCHMARK in godbolt!

[tusooa]

template <typename T>
std::vector<T> to_std(const immer::vector<T>& v)
{
    auto r = std::vector<T>{};
    r.reserve(v.size());
    immer::copy(v, std::back_inserter(v));
    return r;
}

I'm sure now: https://godbolt.org/z/d5n1Wj8MK

it should be immer::copy(v, std::back_inserter(r));

the resulting test result is not that ridiculous, though 1 is still slower.

-------------------------------------------------------------
Benchmark                   Time             CPU   Iterations
-------------------------------------------------------------
benchmark_to_std_1    2517592 ns      1367175 ns          454
benchmark_to_std_2    1232508 ns       527364 ns         1259

[arximboldi]

You are right, sorry!

Corrected godbolt: https://godbolt.org/z/h3cvr7qen