static void *Worker( void *arg ) {
TYPE id = (size_t)arg;
uint64_t entry;
#ifdef FAST
unsigned int cnt = 0, oid = id;
#endif // FAST
unsigned int lid; // local id at each tree level
for ( int r = 0; r < RUNS; r += 1 ) {
entry = 0;
while ( stop == 0 ) {
lid = id; // entry protocol
for ( int lv = 0; lv < depth; lv += 1 ) {
binary_prologue( lid & 1, &t[lv][lid >> 1] );
lid >>= 1; // advance local id for next tree level
} // for
CriticalSection( id );
for ( int lv = depth - 1; lv >= 0; lv -= 1 ) { // exit protocol, retract reverse order
lid = id >> lv;
binary_epilogue( lid & 1, &t[lv][lid >> 1] );
} // for
#ifdef FAST
id = startpoint( cnt ); // different starting point each experiment
cnt = cycleUp( cnt, NoStartPoints );
#endif // FAST
entry += 1;
} // while
#ifdef FAST
id = oid;
#endif // FAST
entries[r][id] = entry;
__sync_fetch_and_add( &Arrived, 1 );
while ( stop != 0 ) Pause();
__sync_fetch_and_add( &Arrived, -1 );
} // for
return NULL;
} // Worker
static void *Worker( void *arg ) {
TYPE id = (size_t)arg;
uint64_t entry;
int other = inv( id ); // int is better than TYPE
#ifdef FAST
unsigned int cnt = 0, oid = id;
#endif // FAST
for ( int r = 0; r < RUNS; r += 1 ) {
entry = 0;
while ( stop == 0 ) {
intents[id] = WantIn; // entry protocol
last = id; // RACE
Fence(); // force store before more loads
if ( FASTPATH( intents[other] != DontWantIn ) ) // local spin
while ( last == id ) Pause(); // busy wait
CriticalSection( id );
intents[id] = DontWantIn; // exit protocol
last = id;
#ifdef FAST
id = startpoint( cnt ); // different starting point each experiment
other = inv( id );
cnt = cycleUp( cnt, NoStartPoints );
#endif // FAST
entry += 1;
} // while
#ifdef FAST
id = oid;
other = inv( id );
#endif // FAST
entries[r][id] = entry;
__sync_fetch_and_add( &Arrived, 1 );
while ( stop != 0 ) Pause();
__sync_fetch_and_add( &Arrived, -1 );
} // for
return NULL;
} // Worker
static void *Worker( void *arg ) {
TYPE id = (size_t)arg;
#ifdef FAST
unsigned int cnt = 0, oid = id;
#endif // FAST
uint64_t entry;
for ( int r = 0; r < RUNS; r += 1 ) {
entry = 0;
while ( stop == 0 ) {
L0: control[id] = WantIn; // entry protocol
Fence(); // force store before more loads
L1: for ( int j = turn; j != id; j = cycleDown( j, N ) )
if ( control[j] != DontWantIn ) { Pause(); goto L1; } // restart search
control[id] = EnterCS;
Fence(); // force store before more loads
for ( int j = N - 1; j >= 0; j -= 1 )
if ( j != id && control[j] == EnterCS ) goto L0;
CriticalSection( id );
// cycle through threads
if ( control[turn] == DontWantIn || turn == id ) // exit protocol
turn = cycleDown( turn, N );
control[id] = DontWantIn;
#ifdef FAST
id = startpoint( cnt ); // different starting point each experiment
cnt = cycleUp( cnt, NoStartPoints );
#endif // FAST
entry += 1;
} // while
#ifdef FAST
id = oid;
#endif // FAST
entries[r][id] = entry;
__sync_fetch_and_add( &Arrived, 1 );
while ( stop != 0 ) Pause();
__sync_fetch_and_add( &Arrived, -1 );
} // for
return NULL;
} // Worker
static void *Worker( void *arg ) {
const TYPE id = (size_t)arg;
uint64_t entry;
#ifdef FAST
unsigned int cnt = 0, oid = id;
#endif // FAST
for ( int r = 0; r < RUNS; r += 1 ) {
entry = 0;
while ( stop == 0 ) {
L0: intents[id] = DontWantIn; // entry protocol
Fence(); // force store before more loads
for ( int j = 0; j < id; j += 1 )
if ( intents[j] == WantIn ) { Pause(); goto L0; }
intents[id] = WantIn;
Fence(); // force store before more loads
for ( int j = 0; j < id; j += 1 )
if ( intents[j] == WantIn ) goto L0;
L1: for ( int j = id + 1; j < N; j += 1 )
if ( intents[j] == WantIn ) { Pause(); goto L1; }
CriticalSection( id ); // critical section
intents[id] = DontWantIn; // exit protocol
#ifdef FAST
id = startpoint( cnt ); // different starting point each experiment
cnt = cycleUp( cnt, NoStartPoints );
#endif // FAST
entry += 1;
} // while
#ifdef FAST
id = oid;
#endif // FAST
entries[r][id] = entry;
__sync_fetch_and_add( &Arrived, 1 );
while ( stop != 0 ) Pause();
__sync_fetch_and_add( &Arrived, -1 );
} // for
return NULL;
} // Worker
static void *Worker( void *arg ) {
TYPE id = (size_t)arg;
uint64_t entry;
int other = inv( id ); // int is better than TYPE
#ifdef FAST
unsigned int cnt = 0, oid = id;
#endif // FAST
for ( int r = 0; r < RUNS; r += 1 ) {
entry = 0;
while ( stop == 0 ) {
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) cc[id] = i % 2; // flicker
#endif // FLICKER
cc[id] = WantIn; // declare intent
Fence(); // force store before more loads
while ( cc[other] == WantIn ) {
if ( FASTPATH( last == id ) ) {
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) cc[id] = i % 2; // flicker
#endif // FLICKER
cc[id] = DontWantIn;
while( cc[other] == WantIn && last == id ) Pause(); // low priority busy wait
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) cc[id] = i % 2; // flicker
#endif // FLICKER
cc[id] = WantIn; // declare intent
Fence(); // force store before more loads
} // if
} // while
CriticalSection( id );
if ( last != id ) {
#ifdef FLICKER
for ( int i = id; i < 100; i += 1 ) last = i % 2; // flicker
#endif // FLICKER
last = id;
} // if
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) cc[id] = i % 2; // flicker
#endif // FLICKER
cc[id] = DontWantIn;
#ifdef FAST
id = startpoint( cnt ); // different starting point each experiment
other = inv( id );
cnt = cycleUp( cnt, NoStartPoints );
#endif // FAST
entry += 1;
} // while
#ifdef FAST
id = oid;
other = inv( id );
#endif // FAST
entries[r][id] = entry;
__sync_fetch_and_add( &Arrived, 1 );
while ( stop != 0 ) Pause();
__sync_fetch_and_add( &Arrived, -1 );
} // for
return NULL;
} // Worker
static void *Worker( void *arg ) {
TYPE id = (size_t)arg;
uint64_t entry;
int other = inv( id ); // int is better than TYPE
#ifdef FAST
unsigned int cnt = 0, oid = id;
#endif // FAST
for ( int r = 0; r < RUNS; r += 1 ) {
entry = 0;
while ( stop == 0 ) {
for ( ;; ) {
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) intents[id] = i % 2; // flicker
#endif // FLICKER
intents[id] = WantIn; // declare intent
// Necessary to prevent the read of intents[other] from floating above the assignment
// intents[id] = WantIn, when the hardware determines the two subscripts are different.
Fence(); // force store before more loads
if ( FASTPATH( intents[other] == DontWantIn ) ) break;
if ( last == id ) {
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) intents[id] = i % 2; // flicker
#endif // FLICKER
intents[id] = DontWantIn;
// Optional fence to prevent LD of "last" from being lifted above store of
// intends[id]=DontWantIn. Because a thread only writes its own id into "last",
// and because of eventual consistency (writes eventually become visible),
// the fence is conservative.
//Fence(); // force store before more loads
await( last != id ); // low priority busy wait
} // if
} // for
CriticalSection( id );
#ifdef FLICKER
for ( int i = id; i < 100; i += 1 ) last = i % 2; // flicker
#endif // FLICKER
last = id; // exit protocol
#ifdef FLICKER
for ( int i = 0; i < 100; i += 1 ) intents[id] = i % 2; // flicker
#endif // FLICKER
intents[id] = DontWantIn;
#ifdef FAST
id = startpoint( cnt ); // different starting point each experiment
other = inv( id );
cnt = cycleUp( cnt, NoStartPoints );
#endif // FAST
entry += 1;
} // while
#ifdef FAST
id = oid;
other = inv( id );
#endif // FAST
entries[r][id] = entry;
__sync_fetch_and_add( &Arrived, 1 );
while ( stop != 0 ) Pause();
__sync_fetch_and_add( &Arrived, -1 );
} // for
return NULL;
} // Worker
