Is an empty aliasing shared_ptr a good alternative to a no-op deleting shared_ptr?

Question

Sometimes I need shared_ptr instances that have a no-op deleter, because an API expects a shared_ptr instance that it wants to store for a limited time but I am given a raw pointer that I am not allowed to own for a time larger than what I am running for.

For this case, I have been using a no-op deleter, such as [](const void *){}, but today I found that there's another alternative to that, using (or abusing?) the aliasing constructor of shared_ptr:

void f(ExpectedClass *ec) {
   std::shared_ptr<ExpectedClass> p(std::shared_ptr<void>(), ec);
   assert(p.use_count() == 0 && p.get() != nullptr);
   apiCall(p);
}

My question is, what is the better way to do this and why? Are the performance expectations the same? With a no-op deleter I expect to pay some cost for the storage of the deleter and reference count, which doesn't appear to be the case when using the aliasing constructor with the empty shared_ptr.

Answer 1

Concerning performance, the following benchmark shows erratic figures:

#include <chrono>
#include <iostream>
#include <limits>
#include <memory>

template <typename... Args>
auto test(Args&&... args) {
    using clock = std::chrono::high_resolution_clock;
    auto best = clock::duration::max();

    for (int outer = 1; outer < 10000; ++outer) {
        auto now = clock::now();

        for (int inner = 1; inner < 20000; ++inner)
            std::shared_ptr<int> sh(std::forward<Args>(args)...);

        auto time = clock::now()-now;
        if (time < best) {
            best = time;
            outer = 1;
        }
    }

    return best.count();
}

int main()
{
    int j;

    std::cout << "With aliasing ctor: " << test(std::shared_ptr<void>(), &j) << '\n'
              << "With empty deleter: " << test(&j, [] (auto) {});
}

Output on my machine with clang++ -march=native -O2:

With aliasing ctor: 11812
With empty deleter: 651502

GCC with identical options gives an even larger ratio, 5921:465794.
And Clang with -stdlib=libc++ yields a whopping 12:613175.

Answer 2

Quick bench with

#include <memory>

static void aliasConstructor(benchmark::State& state) {
  for (auto _ : state) {
    int j = 0;
    std::shared_ptr<int> ptr(std::shared_ptr<void>(), &j);
    benchmark::DoNotOptimize(ptr);
  }
}

BENCHMARK(aliasConstructor);

static void NoOpDestructor(benchmark::State& state) {
  for (auto _ : state) {
    int j = 0;
    std::shared_ptr<int> ptr(&j, [](int*){});
    benchmark::DoNotOptimize(ptr);
  }
}

BENCHMARK(NoOpDestructor);

gives

• a ratio of 1/30 for gcc 10.2
• a ratio of 1/25 for clang 11 libc++

So alias constructor wins.