Emil Axelsson wrote:
Heinrich Apfelmus wrote:
Since every module of my DSL depends on the same global variable, only two things should happen:
* Reloading a module does not reload the global counter. Everything is fine. * Reloading a module does reload the global counter. This forces *all* modules to be reloaded, which removes all expressions that have used the old variable from scope.
But Simon Marlow just said that CAFs are reset upon reload. So the first situation should never occur(?). And as my example shows, resetting the counter does not force all dependent modules to get reloaded.
Something is really weird here. In your example, reloading M2 should not reset the counter, but it does. However, it's not true that all CAFs are reset. After all, this would mean that a would have been reset as well! (Of course, if the two requirements above were met, then it would work.)
However, note that the problem is not so much with Data.Unique; the real problem is that you can't say "I want this to be unique", you have to say "I want this to be unique within this or that *context*". Imagine that I give you two expressions with some Uniques inside, some of them equal, some of them not. How do you know that two equal Uniques denote the same subexpressions? For that, you have to know that I created them in the same context, in the same *environment*.
Using Data.Unique assumes that the environment is one program run. It is clear that reloading modules in GHCi leads to different environments.
OK, so I guess you could say that my solution makes the context explicit by associating it with a time stamp. Of course, this is just an approximation, because it's not impossible for two different contexts to get the same time stamp, especially in a parallel setting. However, it seems that this somewhat ugly trick might actually be what makes this approach to observable sharing work in practice.
Well, what I actually wanted to say is that your values are "free-floating", they contain a lot of "dangling pointers". There is a fundamental problem here which I think you're trying to paste over with judicious use of side effects. Observable sharing is just a funny way of augmenting your DSL with let bindings, the funny thing being that you can use the host-language let instead of an separate Let constructor. I recommend to think it terms of an explicit Let constructor, this will clear up a lot of confusion. For instance, your example can be written as a = Let (1 := "a") `In` Var 1 b = Let (2 := "b") `In` Var 2 (Now, the purpose of Data.Unique is just to supply fresh variable names.) What happens if b is defined as b = Let (1 := "b") `In` Var 1 ? The expressions a and b are both Var 1 , but the variables refer to different environments (1 := "a") and (1 := "b"). If you want to make a large expression from a and b , you have to unify the two environments. For instance, you can use scoping rules (= associated each variable with the height in the expression tree) or you can rename variables. Or you can use unsafePerformIO , which provides one single global environment, eliminating the need to unify environments in the first place. Unfortunately, global environments are unsafe.
Concerning observable sharing, I very much like the approach from
Andy Gill. Type safe observable sharing. http://www.ittc.ku.edu/csdl/fpg/sites/default/files/Gill-09-TypeSafeReificat...
which uses StablePointers . Unfortunately, it forces Typeable contraints on polymorphic combinators.
As far as I've heard, the semantics of stable pointers is not always well-defined either (but I could be wrong). But I should look into whether this technique could be used with the syntactic library. It would be nice to be able to gather several techniques under a common interface.
Err, I meant Stable *Names*, not Stable Pointers, sorry. The worst thing that can happen is that you lose some sharing, but you can never accidentally introduce sharing. Best regards, Heinrich Apfelmus -- http://apfelmus.nfshost.com