Introduce FlattenStrategy.concurrent.
#298
Conversation
andersio
force-pushed
the
concurrent-strategy
branch
from
ec30c09
to
b1cf473
Compare
andersio
force-pushed
the
concurrent-strategy
branch
from
b1cf473
to
23f0b40
Compare
Sources/Flatten.swift
Outdated
| @@ -389,10 +395,11 @@ extension SignalProtocol where Value: SignalProducerProtocol, Error == Value.Err | |||
| } | |||
| } | |||
| } | |||
| state.modify { $0.isStarting = false } | |||
|
|
|||
| isStarting = false | |||
There was a problem hiding this comment.
Is assigning to a Bool an atomic operation? This write could happen concurrently with the read above. I think it needs to at least be in its own Atomic.
I'm not sure that makes sense. Maybe we just add a precondition that count must be >0? |
Sources/Flatten.swift
Outdated
| $0.active = nil | ||
| return !$0.isStarting | ||
| handle.remove() | ||
| _ = cleanup |
There was a problem hiding this comment.
Is this enough to guarantee that producerState remains initialized until after both the read and the write?
There was a problem hiding this comment.
Yes. Its lifetime is limited by the closure or the while loop body scope, whichever is longer.
|
Do we have a way to test for regressions in performance/memory in |
|
Looking at it, I'm not sure why this would cause any performance issues. The code is basically the same. But I'm sure @andersio could run some benchmarks. |
|
It looks like those tests don't compile on Linux. |
Sources/Flatten.swift
Outdated
| $0.active = nil | ||
| return !$0.isStarting | ||
| } | ||
| withExtendedLifetime(deinitializer) { |
There was a problem hiding this comment.
This could probably surround just the producerState.setStarted() below?
Implements #297.
concatandmergeare now aliases ofconcurrent(limit: 1)andconcurrent(limit: .max), respectively. The iterative process to start producers is preserved.WIP: