Benchmark on subprocess with a less trivial example. Run 100x python -c pass: 8.63 sec without my patch, 8.53 sec with my patch => only 1% faster, so the patch is just useless on a real world example.

Finally, I think that there is just nothing to do on Python 2: the overhead between fork(), os.popen and subprocess doesn't impact real world programs. On Python 3: the most critical issue is that close_fds=True is much slower than close_fds=False, but there is already an issue for that (#11284).

So let's close this issue. You can still comment it if you disagree. But if you would like to reopen it: please come with numbers of a benchmark on real programs (not on "exit 0" with a shell or /bin/false).