Unfair Lock, take 2

[andersio]

As mentioned in #282, os_unfair_lock is generally faster all-around (slightly when uncontended, significantly when contended), since it does not need to maintain fairness among waiting threads.

This branch is a transitional implementation purely in Swift, until eventually Swift supports availability propagation of enum cases to switch case scopes.

Related: SR-4079

Lock is an abstract base class of two implementations: UnfairLock and PthreadLock. When constructing a lock through Lock.Type.make, it picks the implementation with #available runtime checks.

It turns out that virtual dispatch (subclassing) is still pretty fast, and is certainly faster than the implementation of the #282 trick in pure Swift, which requires per-instance thunks.

Benchmark Source.

5_000_000 iterations
@Lock: 40.4060875994435 ns
@NSLock: 67.0863552018613 ns
@PthreadLock: 47.6224342023663 ns
@UnfairLock: 40.1545092006554 ns

Existentials are rather slow, by the way.

Inspecting the generated object code with -Owmo reveals that Swift has inlined the vtable. So I guess even if SR-4079 got fixed someday, it would hardly be faster than Lock.

    0x1052c5566 <+38>:   cmpq   0x7bacb(%rip), %rax       ; (void *)0x0000000105350690: type metadata for ReactiveSwift.Lock
    0x1052c556d <+45>:   je     0x1052c58dc               ; <+924> [inlined] function signature specialization <Arg[0] = Dead> of ReactiveSwift.Lock.lock () -> () at Signal.swift:189
    0x1052c5573 <+51>:   cmpq   0x7bac6(%rip), %rax       ; (void *)0x0000000105350738: type metadata for ReactiveSwift.Lock.UnfairLock
    0x1052c557a <+58>:   je     0x1052c558a               ; <+74> [inlined] ReactiveSwift.Lock.UnfairLock.lock () -> () at Signal.swift:302
    0x1052c557c <+60>:   cmpq   0x7bac5(%rip), %rax       ; (void *)0x00000001053507f0: type metadata for ReactiveSwift.Lock.PthreadLock
    0x1052c5583 <+67>:   je     0x1052c559c               ; <+92> [inlined] ReactiveSwift.Lock.PthreadLock.lock () -> () at Signal.swift:302
    0x1052c5585 <+69>:   callq  *0x58(%rax)
    0x1052c5588 <+72>:   jmp    0x1052c55ac               ; <+108> at Signal.swift:303
    0x1052c558a <+74>:   incq   0x8e077(%rip)             ; __profc_/Users/anders/Repo/RAS/Sources/Scheduler.swift:_TFCC13ReactiveSwift13TestSchedulerP33_B7AFF4CC5431FDC431F4B042BAEABD8F15ScheduledActionCfT4dateV10Foundation4Date6actionFT_T__S1_ + 5256
->  0x1052c5591 <+81>:   addq   $0x10, %rdi
    0x1052c5595 <+85>:   callq  0x105329fa2               ; symbol stub for: os_unfair_lock_lock

[liscio]

This all looks good to me.

[sharplet]

It's a bummer that this uses both #if os and @available :(