On Mon, Oct 22, 2012 at 2:17 AM, Clément Bera clement.bera@inria.fr wrote:
Hello,
I don't understand something on BlockClosure>>ensure:. Why does it use 'self valueNoContextSwitch' and not 'self value' ? In which case is there an issue ?
Prior to my closure implementation BlockContext>>value[:value:*] was not a context switch point. So to preserve the threading semantics of ensure: we implemented ensure: in terms of BlockClosure>>valueNoContextSwitch, which is not a context-switch point, as opposed to BlockClosure>>value[:value:*]. The context switch points in the VM are non-primitive sends (or failing primitives, i.e. full method activation), backward branches, the Process primitives suspend, resume, signal and wait (and CrtiticalSection/Mutex enter and exit) and block evaluation.
Thank you for any answers
BlockClosure>>ensure: is implemented this way :
ensure: aBlock "Evaluate a termination block after evaluating the receiver, regardless of whether the receiver's evaluation completes. N.B. This method is *not* implemented as a primitive. Primitive 198 always fails. The VM uses prim 198 in a context's method as the mark for an ensure:/ifCurtailed: activation."
| complete returnValue | <primitive: 198> returnValue := *self valueNoContextSwitch*. complete ifNil:[ complete := true. aBlock value. ]. ^ returnValue
On 24 October 2012 00:57, Eliot Miranda eliot.miranda@gmail.com wrote:
On Mon, Oct 22, 2012 at 2:17 AM, Clément Bera clement.bera@inria.fr wrote:
Hello,
I don't understand something on BlockClosure>>ensure:. Why does it use 'self valueNoContextSwitch' and not 'self value' ? In which case is there an issue ?
Prior to my closure implementation BlockContext>>value[:value:*] was not a context switch point. So to preserve the threading semantics of ensure: we implemented ensure: in terms of BlockClosure>>valueNoContextSwitch, which is not a context-switch point, as opposed to BlockClosure>>value[:value:*]. The context switch points in the VM are non-primitive sends (or failing primitives, i.e. full method activation), backward branches, the Process primitives suspend, resume, signal and wait (and CrtiticalSection/Mutex enter and exit) and block evaluation.
Hmm.. your response puzzles me even more.
just simple example to show a controversy:
[ do something ] ensure: [ semaphore signal. anotherSemaphore wait ].
I think that semantics of ensure should be: - same as #value, but evaluate additional "ensure" block prior return point , be it normal return or stack unwinding (non-local return or process termination).
Anything else, like scheduler interference, which may preempt current process which evaluates block doesn't matters because semantics is local to process scope where given block is evaluated.
Trying to ensure "ensure" semantics even regardless of scheduling is pointless and futile. So, i am puzzled, what exactly "threading semantics" you trying to preserve here.
Igor Stasenko wrote
Hmm.. your response puzzles me even more.
just simple example to show a controversy:
[ do something ] ensure: [ semaphore signal. anotherSemaphore wait ].
That isn't the point Eliot is making. The #valueNoContextSwitch affects the *receiver* of the ensure: message (not the argument) and only insofar as activating (entering) it is concerned; i.e., it prevents a process switch while entering the ensured block. I believe that originally we were trying to avoid any deviation from the interpreter's behavior and added this simply to ensure complete compatibility with the context VMs, but in hindsight I'm not certain this ever had any effect.
Cheers, - Andreas
-- View this message in context: http://forum.world.st/Re-Pharo-project-BlockClosure-ensure-implementation-tp... Sent from the Squeak VM mailing list archive at Nabble.com.
vm-dev@lists.squeakfoundation.org