cache bug

st to prove an optinmization in function get_singleton has bug// dew to the cache-unconscious programming// Ekin for Vladimir Chebotarev// Moscow, 2012// no rights reserved// run: g++ test.cpp -lpthread && ./a.out 16 10 0.005#include <stdlib.h>#include <pthread.h>#include <time.h>#include <stdio.h>#include <math.h>#define THREAD_MAX 1024#define CPU_NUM  16#define OOPS(A)  do { \    fprintf(stderr, "oops: local = %p;" #A " = %p; thread #%d\n" \            , t->local, A, t->num); \    exit(__LINE__); } while(0)struct tls {    void*    local; // thread's view of singleton    int      num;    int      iter_count;    pthread_t thread;    cpu_set_t cpu_set;    timespec pause_time_requested;    timespec pause_time_remaining;     double busy; };struct tls_elem {    tls  data;    char pad[0x1000 - sizeof(data)]; };tls_elem* tlses = 0;pthread_mutex_t mutex;pthread_barrier_t barrier;static void* volatile singleton = 0; // master copy of singletonvoid* get_singleton() {    if (!singleton) {        pthread_mutex_lock(&mutex);        if (!singleton)            singleton = malloc(sizeof(int));        pthread_mutex_unlock(&mutex); }    return singleton; }void* run(void* obj) {    tls* t = static_cast<tls*>(obj);    while(t->iter_count--) {        //fprintf(stderr, "thread #%4d started  iter = #%d\n", t->num, t->iter_count);        //nanosleep(&t->pause_time_requested, &t->pause_time_remaining);        for (int i  = 0; i < 10000; ++i)            t->busy += sin((double)i);        pthread_barrier_wait(&barrier);        void* gotten = get_singleton();        if (!t->local)            t->local = gotten;        else if (t->local != gotten)            OOPS(gotten);        //fprintf(stderr, "thread #%4d finished  iter = #%d\n", t->num, t->iter_count);        pthread_barrier_wait(&barrier); }    return obj; }int main (int argc, char* argv[]) {    if (argc != 4) {        fprintf(stderr, "USAGE: %s thread_count iter_count pause_time\n"                , argv[0]);        exit(__LINE__); }    int thread_count = atoi(argv[1]);    int iter_count = atoi(argv[2]);    double pause_time = atof(argv[3]);    time_t pause_sec = (time_t)pause_time;    long   pause_nanosec = (long)(1e9*(pause_time - pause_sec));    if (!(thread_count && iter_count && thread_count < THREAD_MAX)) {        fprintf(stderr, "INVALID ARGUMENT\n");        exit(__LINE__); }    pthread_mutex_init(&mutex, 0);    pthread_barrier_init(&barrier, 0, thread_count + 1);        // "+1" because of the main thread will wait too    tlses = static_cast<tls_elem*>(malloc(thread_count * sizeof(tls_elem)));    fprintf(stderr, "TEST STARTED\n");    int i;    for (i = 0; i < thread_count; ++i) {        tls* t = &(tlses + i)->data;        t->local = 0;        t->num = i;        CPU_ZERO(&t->cpu_set);        CPU_SET(i % CPU_NUM, &t->cpu_set);        t->iter_count = iter_count;        t->pause_time_requested.tv_sec  = pause_sec;        t->pause_time_requested.tv_nsec = pause_nanosec;        t->busy = 0.;        pthread_create(&t->thread, 0, run, t);         pthread_setaffinity_np(t->thread, sizeof(t->cpu_set), &t->cpu_set); }    while(iter_count--) {        fprintf(stderr, "iter #%d started\n", iter_count);        pthread_barrier_wait(&barrier);        pthread_barrier_wait(&barrier);        //fprintf(stderr, "iter #%d finished\n", iter_count);        for (i = 0; i < thread_count; ++i) {            tls* t = &(tlses + i)->data;            if (t->local != singleton || !t->local)                OOPS(singleton);            t->local = 0; }        free(singleton);        singleton = 0;        tls* t = &tlses->data;        //fprintf(stderr, " main thread pause iter #%d\n", iter_count);        nanosleep(&t->pause_time_requested, &t->pause_time_remaining);}    fprintf(stderr, "SUCCESS (no bug)\n");    for (i = 0; i < thread_count; ++i)        pthread_join((tlses + i)->data.thread, 0);    free(tlses);    return 0; }