CODEWITHSUNDEEP
gccoptimizationboostclangc++14
zschoche
zschoche

Reputation: 41

boost::variant vs. polymorphism, very different performance results with clang and gcc

I'm trying to figure out how much the execution time of boost::variant differ from a polymorphism approach. In my first test I got very different results on gcc 4.9.1 and clang+llvm 3.5.

You can find the code below. Here are my results:

clang+llvm

polymorphism: 2.16401
boost::variant: 3.83487

gcc:

polymorphism: 2.46161
boost::variant: 1.33326

I compiled both with -O3. Is someone able to explain that?

code

#include <iostream>
#include <vector>
#include <algorithm>
#include <boost/variant.hpp>
#include <boost/variant/apply_visitor.hpp>
#include <ctime>

struct value_type {

    value_type() {}
    virtual ~value_type() {}
    virtual void inc() = 0;
};

struct int_type : value_type {

    int_type() : value_type() {}
    virtual ~int_type() {}

    void inc() { value += 1; }

      private:
    int value = 0;
};

struct float_type : value_type {

    float_type() : value_type() {}
    virtual ~float_type() {}

    void inc() { value += 1; }

      private:
    float value = 0;
};

void dyn_test() {

    std::vector<std::unique_ptr<value_type>> v;
    for (int i = 0; i < 1024; i++) {
        if (i % 2 == 0)
            v.emplace_back(new int_type());
        else
            v.emplace_back(new float_type());
    }

    for (int i = 0; i < 900000; i++) {
        std::for_each(v.begin(), v.end(), [](auto &item) { item->inc(); });
    }
}

struct visitor : boost::static_visitor<> {

    template <typename T> void operator()(T &item) { item += 1; }
};

using mytype = boost::variant<int, float>;
void static_test() {
    std::vector<mytype> v;
    for (int i = 0; i < 1024; i++) {
        if (i % 2 == 0)
            v.emplace_back(0);
        else
            v.emplace_back(0.f);
    }

    visitor vi;
    for (int i = 0; i < 900000; i++) {
        std::for_each(v.begin(), v.end(), boost::apply_visitor(vi));
    }
}

template <typename F> double measure(F f) {
    clock_t start = clock();
    f();
    clock_t end = clock();
    float seconds = (float)(end - start) / CLOCKS_PER_SEC;
    return seconds;
}

int main() {
    std::cout << "polymorphism: " << measure([] { dyn_test(); }) << std::endl;
    std::cout << "boost::variant: " << measure([] { static_test(); }) << std::endl;

    return 0;
}

assembler

gcc

clang+llvm

Upvotes: 4

Views: 1844

Answers (1)

Puppy
Puppy

Reputation: 147028

Clang is known to miscompile some std::vector functions from various Standard libraries, due to some edge cases in their inliner. I don't know if those have been fixed by now but quite possibly not. Since unique_ptr is smaller and simpler than boost::variant it's more likely that it does not trigger these edge cases.

The code you post is practically "Why boost::variant is great". A dynamic allocation and random pointer index in addition to the regular indirections that both perform? That's a heavy hit (relatively).

Upvotes: 2

Related Questions