On 22 mai 2014, at 22:29, Luc Fabresse <luc.fabresse@gmail.com> wrote:
2014-05-22 19:49 GMT+02:00 Camille Teruel <camille.teruel@gmail.com>:
On 22 mai 2014, at 19:07, Luc Fabresse <luc.fabresse@gmail.com> wrote:
2014-05-22 18:06 GMT+02:00 Camille Teruel <camille.teruel@gmail.com>:
On 22 mai 2014, at 11:54, stepharo <stepharo@free.fr> wrote:
In general, we should think about if is would make sense that the language comes with a good proxy model by default.
+ 1 We should do that with Camille.
Yes. Now the problem is to agree on what is a "good model". There is many use cases for proxies and they come with their own models and problems.
yes sure
Here are some requirements I can think of:
- Uniformity: all objects can be wrapped, even special ones (it's a technicality but it's still important) - Interception of all message: currently we cannot intercept #== and optimized control-flow messages like #ifTrue:ifFalse #to:do:, #whileTrue:, etc... - Encapsulation (a.k.a the self problem): avoid leaking of references to the target. E.g. "aProxy yourself" should return the proxy not the target. - Behavioral intercession: Not everything is a message send. What other events do we want to intercept? IV access is a must-have for me. - Composition: Wrapping a proxy with another one should work as expected, i.e. applying both policies in order. - Performances: intercepting message sends is slow, intercepting iv accesses is even worse (relying on code instrumentation). We would need VM support. - Security: It shouldn't be possible to by-pass or corrupt a proxy policy. - Delimited replacement: Ensure that a proxy is always used in place of its target throughout the processing of a message send, no matter the alias used (that is not necessarily "self"). Think about a read-only proxy for example.
Right now, to intercept messages we have to rely on #cannotInterpret: hack
why do you call it a hack?
Because it's a hijack a safety check of the VM. We set the method dictionary of the proxy's class to nil (or of its superclass if the proxy must understand some messages). Since the VM doesn't know what to do with that, it calls #cannotInterpret: starting the lookup in the superclass of that class. This superclass then redefines #cannotInterpret: to hand over the message interception to the proxy's handler...
I know the mechanism ;-)
I know you know, but there is many people on this mailing list :). It was to stress that we need three indirections (1st lookup, 2nd lookup, forward request to handler) where 1 would be enough if done at the vm side. I call it a hack but it's not pejorative. It's the best we can do without vm modification after all.
but to me it is not hack it is a clever idea that let the programmer to decide what to do at the image side to implement crazy stuff without modifying the VM
That sure is clever and it's great to not have to modify the vm, but modifying it to support proxies worth it in my opinion.
The VM is simple (one interception hook) while it enables to capture a lot of things. Of course we can imagine something else at the VM side but the more code in the VM the less flexible it will be
do you consider DNU as hack too?
Using it to implement proxies? Yes. But in addition, DNU is bad since you can only intercept messages that are not understood. And you can also by-pass the message the proxy does understand by sending directly #doesNotUnderstand: to the proxy with a specific message as argument.
we also explained that in the Ghost paper... that is why cannotInterpret is better but yet incompelete to capture everything
Only because of compilation optimizations. It is a real problem for certain applications.
. To intercept iv access or other events we have to rely on code instrumentation and delegation.
yes but I would like to investigate if cannot achieve iv interception using special slots
I doubt we can since it is the target's slots that are taken into account during compilation. Or else every slot in the system must be special => really poor performances.
I have no clear idea yet on that so I must have look and you are certainly right
Luc
It would be easier if the VM would allow to create an object that controls its own interpretation (via a metaobject that defines a MOP for example). Such an object should at least be able to specify its message lookup and how to reads and writes IVs. And optionally: - how it sends messages this one is possible with cannotInterpret you intercept and then you decide what to do and maybe execute something else so you control message sending
I was talking about interception of message sending (the message the target would have sent) not of interception of message reception (the living wage of any proxy implementation :) ). Like the others events, it also lies at the submethod level.
- how it resolves other variables (temporaries, arguments, globals, class variables) - how it resolves literals - how it executes primitives (because else the VM may cry) - how it is assigned to a variable or passed in a message (to control static/dynamic aliasing respectively)
all of those ones are submethod level and yes Ghost stops at the object granularity currently
- ... many others ... With such objects (+ operating by delegation instead of forwarding) we can solve most of the requirements listed above (well except performant delimited replacement). All that to say that it is not easy...
yes but it is interesting ;-) and we also discuss "some parts" of that with Nick
Great, let us know about the conclusions :)
Luc
(this way we would avoid that everyone creates one that is not completely working)