main.cpp

extern void run();

int main(int argc, char* argv[]) {
  run();
}

Makefile

.PHONY: main

main: main.cpp libthread_local.so
	g++ -std=c++11 -O2 main.cpp -L . -lthread_local -o main

libthread_local.so: thread_local.cpp
	g++ -std=c++11 -O2 -fPIC -shared -o libthread_local.so thread_local.cpp

.PHONY: clear
clean:
	rm -f main libthread_local.so

thread_local.cpp

#include <chrono>
#include <cmath>
#include <cstdlib>
#include <cstring>
#include <time.h>
#include <iostream>
#include <atomic>

double diff(timespec start, timespec end)
{
  timespec temp;
  if ((end.tv_nsec-start.tv_nsec)<0) {
    temp.tv_sec = end.tv_sec-start.tv_sec-1;
    temp.tv_nsec = 1000000000+end.tv_nsec-start.tv_nsec;
  } else {
    temp.tv_sec = end.tv_sec-start.tv_sec;
    temp.tv_nsec = end.tv_nsec-start.tv_nsec;
  }
  return (double)temp.tv_sec + temp.tv_nsec * 1e-9;
}

template <typename Op>
double benchmark(Op op) {
  const int N = 10000000;
  timespec time1, time2;
  op();
  clock_gettime(CLOCK_MONOTONIC, &time1);
  for (int i = 0; i < N; i++)
    op();
  clock_gettime(CLOCK_MONOTONIC, &time2);
  double time = diff(time1, time2);
  std::cout << (time/N)*1e9 << " ns \n";
  return time;
}

__attribute__((noinline)) 
void add_thread_local() {
	static thread_local int x;
	x += 1;
}

__attribute__((noinline)) 
void add_global() {
	static int x;
	x += 1;
}

void run() {
  benchmark([&]() {
    add_thread_local();
  });
  benchmark([&]() {
    add_global();
  });
}

Results below. The difference between the two numbers approximates the "cost" of thread-locals access in shared libraries (about 1-1.5 ns).

Linux (Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz):

3.61949 ns
1.9507 ns

macOS (2.4 GHz 8-Core Intel Core i9)

2.3994 ns
1.3577 ns