> That sounds like a micro-optimization which is not worth it.

In the back of my mind I was also thinking to generalize the behavior at runtime whenever a signal is tripped by a non-Python thread (e.g. a thread created by an extension module or ctypes), instead of special-casing Windows. 

To examine this, I created a C library in Linux that defines a test() function that creates two threads via pthread_create(). The first thread sleeps for 10 seconds, and the second thread sleeps for 5 seconds and then calls pthread_kill() to send a SIGINT to the first thread. In 3.8, calling the test() function via ctypes followed by executing an infinite loop will interrupt the loop with a KeyboardInterrupt as soon as the second thread sends SIGINT. But in 3.10, the loop never gets interrupted because the C signal handler isn't called on the main thread, so eval_breaker never gets set.