static inline bool trylock( TYPE id ) { // based on Lamport-Fast algorithm
b[id] = true;
x = id;
Fence(); // force store before more loads
if ( FASTPATH( y != N ) ) {
b[id] = false;
return false;
} // if
y = id;
Fence(); // force store before more loads
if ( FASTPATH( x != id ) ) {
b[id] = false;
Fence(); // OPTIONAL, force store before more loads
for ( int k = 0; k < N; k += 1 )
await( ! b[k] );
if ( FASTPATH( y != id ) ) return false;
} // if
bool leader;
if ( ! fast ) { fast = true; leader = true; }
else leader = false;
y = N;
b[id] = false;
return leader;
} // WCas
static inline bool trylock( TYPE id ) { // based on Burns-Lamport algorithm
b[id] = true;
Fence(); // force store before more loads
for ( typeof(id) thr = 0; thr < id; thr += 1 ) {
if ( FASTPATH( b[thr] ) ) {
b[id] = false;
return false ;
} // if
} // for
for ( typeof(id) thr = id + 1; thr < N; thr += 1 ) {
await( ! b[thr] );
} // for
bool leader;
if ( ! fast ) { fast = true; leader = true; }
else leader = false;
b[id] = false;
return leader;
} // WCas
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;
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
