Comparing __thread storage class specifier to pthread_getspecific on OS X 10.7+

tls-vs-pthread_getspecific.cc

// Benchmarking pthread_getspecific vs __thread on OS X
//
// Compile with clang++ -g -O3 -std=c++11

#include <stdio.h>
#include <pthread.h>

#include <chrono>
#include <cinttypes>

__thread int64_t foo;

inline int64_t thread_specifier() {
    return foo;
}

pthread_key_t key;

inline int64_t pthread_getspecific() {
    return reinterpret_cast<int64_t>(pthread_getspecific(key));
}

template<typename Func>
std::chrono::microseconds run(Func const& func) {
    auto start = std::chrono::system_clock::now();
    func();
    auto end = std::chrono::system_clock::now();
    return std::chrono::duration_cast<std::chrono::microseconds>(end - start);
}

int main(int argc, char **) {
    const int kIters = 1E7;
    pthread_key_create(&key, nullptr);
    foo = 0;
    int64_t val = 0;

    auto t1 = run([&val]() -> void {
            int ret = 0;
            #pragma nounroll
            for (int i = 0; i < kIters; ++i) {
                ret += pthread_getspecific();
                asm volatile("" : "+r" (ret));
            }
            val += ret;
        });

    auto t2 = run([&val]() -> void {
            int ret = 0;
            #pragma nounroll
            for (int i = 0; i < kIters; ++i) {
                ret += thread_specifier();
                asm volatile("" : "+r" (ret));
            }
            val += ret;
        });

    printf("us/iter, pthread: %f ts: %f\n", t1.count() / (double) kIters,
           t2.count() / (double) kIters);

    return 0;
}