[Vm-dev] Interpreter versus StackInterpreter hierarchy
Ben Coman
btc at openinworld.com
Sat May 21 02:52:13 UTC 2016
On Sat, May 21, 2016 at 4:36 AM, Clément Bera <bera.clement at gmail.com> wrote:
> On Fri, May 20, 2016 at 7:51 PM, Ben Coman <btc at openinworld.com> wrote:
>> On Fri, May 20, 2016 at 9:25 PM, Clément Bera <bera.clement at gmail.com> wrote:
>> > On Thu, May 19, 2016 at 3:42 PM, Ben Coman <btc at openinworld.com> wrote:
>> >> P.S. For the curious, here is the proof of concept I tried. A single
>> >> call to the primitive counts up to four by backing up the
>> >> instructionPointer such that the primitive is executed again, until
>> >> the exit condition of four is reached.
>> >>
>> >> (Note the use of Semaphore and ExcessSignalsIndex is not significant,
>> >> just an expedient template I was familiar with.)
>> >>
>> >> # VM SIDE...
>> >>
>> >> StackInterpreterPrimitives >> primitiveRetryExperiment
>> >> | excessSignals stackTop |
>> >> stackTop := self stackTop.
>> >> excessSignals := objectMemory fetchInteger: ExcessSignalsIndex
>> >> ofObject: stackTop.
>> >> excessSignals := excessSignals + 1.
>> >> objectMemory storeInteger: ExcessSignalsIndex
>> >> ofObject: stackTop
>> >> withValue: excessSignals.
>> >> [ excessSignals > 3 ] ifFalse: [ instructionPointer :=
>> >> instructionPointer - 1 ].
>> >>
>> >> StackInterpreter class >> initializePrimitiveTable
>> >> (234 primitiveRetryExperiment)
>> >>
>> >
>> > ... Why instructionPointer - 1 ?
>>
>> A random experiment before moving on to -2, -3 and -4. Obviously I
>> don't know enough yet to have properly judged its impact. Just poking
>> it with a stick to see what pops out, and following that old adage
>> that if you need something to be true, "assume it" until you learn
>> otherwise. And whadayaknow..., it worked for the StackInterpreter.
>> And I've since learnt its not so easy for the JIT.
>>
>> > It works if the send is encoded in a single byte as by chance is the case in your example, else your interpretation get misaligned, which is a complete nonsense (unless you're working with a Smalltalk-78 VM ;-) ). Going backward in the instructions is not that trivial, you need a scanner to find the previous pc.
>>
>> I thought *maybe* that since the layout of the receiver is "known" (to
>> be an OwnedLock) the previous pc could be assumed at a known distance
>> up the stack - but maybe that is bad practice and trouble if the
>> layout changes for anyone subclassing OwnedLock.
>
> I don' really understand that known layout of the receiver thing.
Just my limited understanding anf naive assumption. The course I took
on compilers was 20 years ago. You counter example below helps...
> The problem is with the encoding of sends in the bytecode. They are encoded in different number of bytes to lower the memory footprint.
>
> For example, in Pharo, in this method:
> foo
> PrimExp new primRetryExperiment excessSignals
>
> => 27 <D1> send: primRetryExperiment
>
> the send to primRetryExperiment is encoded in a single byte.
>
> But in this method:
> foo
> self foo1 foo2 foo3 foo4 foo5 foo6 foo7 foo8
> self foo9 foo10 foo11 foo12 foo13 foo14 foo15 foo16.
> PrimExp new primRetryExperiment excessSignals
>
> => 109 <83 11> send: primRetryExperiment
>
> The same send is encoded in 2 bytes.
> So, after the send, in the primitive code, you need the instruction pointer to go back by 1 or 2 bytes to repeat the send. But you don't know by how many bytes. There is no easy way to guess if it's 1 or 2 (or something else, there are other cases I omitted).
Okay. That helps me understand of my task better.
>> > See #skipBackBeforeJump for example, which goes backward 1 instruction in the image side. You need to implement something similar in the interpreter if you want to do that... but you don't.
>>
>> I'll take a look just for interest, but point taken.
>>
>> > Although going back one instruction is very funny, it won't work with the JIT, unless you do something completely evil, crazy and hackish.
>>
>> That was the advice I was looking for. I don't want to be hackish ;)
>> At least not in the final result.
>>
>> > Most likely you want in fact to implement your primitive with a while loop instead, so you don't need that ip modification hack. Why didn't you do a while loop in the first place ?
>>
>> A loop in the Image would work, but I'm attempting to keep it in the
>> VM. IIUC a loop won't work in the VM because the primitive sleeps and
>> changes context if the lock is held by someone else.
>
>
> hum. If I change your example to:
>
> StackInterpreterPrimitives >> primitiveRetryExperiment
> | excessSignals stackTop |
> stackTop := self stackTop.
> excessSignals := objectMemory fetchInteger: ExcessSignalsIndex ofObject: stackTop.
> [ excessSignals > 3 ] whileFalse: [
> excessSignals := excessSignals + 1.
> objectMemory storeInteger: ExcessSignalsIndex
> ofObject: stackTop
> withValue: excessSignals ].
>
> It's exactly equivalent to what you wrote, isn't it ? There is no process switch - context switch - in both cases in the loop, isn't it ? Or maybe I am missing something ?
Apologies for mis-leading the discussion. The missing bit is, the
above isn't my actually requirement. Its only an experiment of the
effect of modifying the instruction pointer. My real use case is for
the OwnedLock primitives I'm introducing [1], which at the end of
primitiveOwnedLockWaitAcquire.
self transferTo: self wakeHighestPriority from: CSOwnedLockWaitAcquire.
self forProcessPrimitiveReturnToExecutivePostContextSwitch: inInterpreter.
That code works well, except for one corner case identified by Dennis
where after primitiveOwnedLockRelease hands the lock to a waiting
process, that process could be terminated before waking up. Doing
that in primitiveOwnedLockRelease is required because when
primitiveOwnedLockWaitAcquire goes to sleep, you've already past its
code that acquires the lock.
Mutex>>critical: mutuallyExcludedBlock
^[ lock waitAcquire
ifNil: mutuallyExcludedBlock
ifNotNil: mutuallyExcludedBlock
] ensure: [lock release].
instead of the faster...
Mutex>>critical: mutuallyExcludedBlock
^lock waitAcquire
ifNil: mutuallyExcludedBlock
ifNotNil:[ mutuallyExcludedBlock ensure: [lock release] ].
I'd heard about Eliot "retrying primitives" for forwarding failures
and considered a solution might be for primitiveOwnedLockWaitAcquire
to be re-executed when its process woke up. So the lock is only
acquired after the process wakes up, rather than being handed to it
while it was asleep. Conceptually the process sleeps at the top of
the primitive rather than the bottom.
I found what Eliot does...
dispatchFunctionPointer(primitiveFunctionPointer);
/* begin maybeRetryFailureDueToForwarding */
if (GIV(primFailCode)
&& (checkForAndFollowForwardedPrimitiveState())) {
/* begin initPrimCall */
GIV(primFailCode) = 0;
dispatchFunctionPointer(primitiveFunctionPointer);
}
but I guess that won't work when switching processes. So having
caught the idea for primitive retry, I went looking for another way
that worked across a process switch, and learn a bit more about the VM
in the process. Hence my experiment to naively decrement the
instruction pointer before switching context.
[1] http://forum.world.st/OwnedLock-primitives-request-for-review-td4886130.html
cheers -ben
>>
>> > Primitive calls are not interrupt points anyway, it would be exactly the same behavior, wouldn't it ?
>> >
>> >>
>> >>
>> >> # IMAGE SIDE...
>> >>
>> >> Semaphore subclass: #PrimExp
>> >> instanceVariableNames: ''
>> >> classVariableNames: ''
>> >> package: '0PrimitiveRetryExperiment'
>> >>
>> >> PrimExp >> initialize
>> >> excessSignals := 1.
>> >>
>> >> PrimExp >> primRetryExperiment
>> >> <primitive: 234>
>> >>
>> >> PrimExp >> excessSignals
>> >> ^ excessSignals
>> >>
>> >> # TEST CASE...
>> >>
>> >> PrimExp new primRetryExperiment excessSignals
>> >> --> 4
>> >>
>> >> I've only done this with the Stack VM so far. I'll report further
>> >> when I try it with Cog.
>>
>> So I found the approach doesn't work with Cog.
>> cheers -ben
>>
>> >
>> > It's nice to see people trying to hack the VM :-). I'll try to answer your other questions.
More information about the Vm-dev
mailing list