Why does using std::endl with ostringstream affect output speed?

Question

I'm timing the difference between various ways to print text to standard output. I'm testing cout, printf, and ostringstream using both \n and std::endl. I expected std::endl to make a difference with cout (and it did), but I didn't expect it to slow down output with ostringstream. I thought using std::endl would just write a \n to the stream and it would still only get flushed once. What's going on here? Here's all my code:

// cout.cpp
#include <iostream>

using namespace std;

int main() {
  for (int i = 0; i < 10000000; i++) {
    cout << "Hello World!\n";
  }
  return 0;
}

// printf.cpp
#include <stdio.h>

int main() {
  for (int i = 0; i < 10000000; i++) {
    printf("Hello World!\n");
  }
  return 0;
}

// stream.cpp
#include <iostream>
#include <sstream>

using namespace std;

int main () {
  ostringstream ss;
  for (int i = 0; i < 10000000; i++) {
    ss << "stream" << endl;
  }
  cout << ss.str();
}

// streamn.cpp
#include <iostream>
#include <sstream>

using namespace std;

int main () {
  ostringstream ss;
  for (int i = 0; i < 10000000; i++) {
    ss << "stream\n";
  }
  cout << ss.str();
}

And here's my Makefile

SHELL:=/bin/bash

all: cout.cpp printf.cpp
    g++ cout.cpp -o cout.out
    g++ printf.cpp -o printf.out
    g++ stream.cpp -o stream.out
    g++ streamn.cpp -o streamn.out
time:
    time ./cout.out > output.txt
    time ./printf.out > output.txt
    time ./stream.out > output.txt
    time ./streamn.out > output.txt

Here's what I get when I run make followed by make time

time ./cout.out > output.txt

real    0m1.771s
user    0m0.616s
sys 0m0.148s
time ./printf.out > output.txt

real    0m2.411s
user    0m0.392s
sys 0m0.172s
time ./stream.out > output.txt

real    0m2.048s
user    0m0.632s
sys 0m0.220s
time ./streamn.out > output.txt

real    0m1.742s
user    0m0.404s
sys 0m0.200s

These results are consistent.

Answer 1

std::endl triggers a flush of the stream, which slows down printing a lot. See http://en.cppreference.com/w/cpp/io/manip/endl

It is often recommended to not use std::endl unless you really want the stream to be flushed. If this is really important to you, depends on your use case.

Regarding why flush has a performance impact even on a ostringstream (where no flushing should happen): It seems that an implementation is required to at least construct the sentry objects. Those need to check good and tie of the ostream. The call to pubsync should be able to be optimized out. This is based on my reading of libcpp and libstdc++.

After some more reading the interesting question seems to be this: Is an implementation of basic_ostringstream::flush really required to construct the sentry object? If not, this seems like a "quality of implementation" issues to me. But I actually think it needs to because even a basic_stringbug can change to have its badbit set.