Am 31.01.2014 10:18, schrieb Stanislav Blinov:
You forgot to make the flag static for AtomicSingleton. I'd also move
the timing into the threads themselves, for fairness :)
http://codepad.org/gvm3A88k
Timings on my machine:
ldc2 -unittest -release -O3:
Test time for LockSingleton: 537 msecs.
Test time for SyncSingleton: 2 msecs.
Test time for AtomicSingleton: 2.25 msecs.
dmd -unittest -release -O -inline:
Test time for LockSingleton: 451.5 msecs.
Test time for SyncSingleton: 7.75 msecs.
Test time for AtomicSingleton: 99.75 msecs.
For x86 CPUs you don't really need MemoryOrder.acq as reads are atomic
by default. So I replaced that with MemoryOrder.raw and named it
AtomicSingletonRaw
On Windows 7:
dmd -unittest -release -O -inline -noboundscheck
Test time for LockSingleton: 299 msecs.
Test time for SyncSingleton: 5 msecs.
Test time for AtomicSingleton: 304 msecs.
Test time for AtomicSingletonRaw: 280 msecs.
ldc2 -release -unittest -O3
Test time for LockSingleton: 320 msecs.
Test time for SyncSingleton: 2 msecs.
Test time for AtomicSingleton: 271 msecs.
Test time for AtomicSingletonRaw: 209 msecs.
It seems that the SyncSingleton is supperior in all cases.
