Peterson's algorithm built from plain non-atomic variables

Peterson's algorithm is the textbook proof that two threads can achieve mutual exclusion using only plain reads and writes of shared flag[]/turn, with no atomic read-modify-write instruction required. Compiled and run as plain (non-atomic, non-volatile) C++ variables on a real modern multicore machine with a standard optimizing compiler, is mutual exclusion actually guaranteed?

main.cpp
bool flag[2] = {false, false};
int turn = 0;

void thread0() {
    flag[0] = true;
    turn = 1;
    while (flag[1] && turn == 1) { /* spin */ }
    // critical section
    flag[0] = false;
}

void thread1() {
    flag[1] = true;
    turn = 0;
    while (flag[0] && turn == 0) { /* spin */ }
    // critical section
    flag[1] = false;
}

26 of 51 in Concurrency