MLtonContIsolateImplementation

As noted before, it is fairly easy to get the operational behavior of isolate with just callcc and throw, but establishing the right space behavior is trickier. Here, we show how to start from the obvious, but inefficient, implementation of isolate using only callcc and throw, and derive an equivalent, but more efficient, implementation of isolate using MLton’s primitive stack capture and copy operations. This isn’t a formal derivation, as we are not formally showing the equivalence of the programs (though I believe that they are all equivalent, modulo the space behavior).

Here is a direct implementation of isolate using only callcc and throw:

We use the standard nested callcc trick to return a continuation that is ready to receive an argument, execute the isolated function, and exit the program. Both Exit.topLevelSuffix and MLtonExn.topLevelHandler will terminate the program.

Throwing to an isolated function will execute the function in a semantically empty context, in the sense that we never re-execute the original continuation of the call to isolate (i.e., the context that was in place at the time isolate was called). However, we assume that the compiler isn’t able to recognize that the original continuation is unused; for example, while we (the programmer) know that Exit.topLevelSuffix and MLtonExn.topLevelHandler will terminate the program, the compiler may only see opaque calls to unknown foreign-functions. So, that original continuation (in its entirety) is part of the continuation returned by isolate and throwing to the continuation returned by isolate will execute f x (with the exit wrapper) in the context of that original continuation. Thus, the garbage collector will retain everything reachable from that original continuation during the evaluation of f x, even though it is semantically garbage.

Note that this space-leak is independent of the implementation of continuations (it arises in both MLton’s stack copying implementation of continuations and would arise in SML/NJ’s CPS-translation implementation); we are only assuming that the implementation can’t see the program termination, and so must retain the original continuation (and anything reachable from it).

So, we need an empty continuation in which to execute f x. (No surprise there, as that is the written description of isolate.) To do this, we capture a top-level continuation and throw to that in order to execute f x:

We presume that base is evaluated early in the program. There is a subtlety here, because one needs to believe that this base continuation (which technically corresponds to the entire rest of the program evaluation) works as an empty context; in particular, we want it to be the case that executing f x in the base context retains less space than executing f x in the context in place at the call to isolate (as occurred in the previous implementation of isolate). This isn’t particularly easy to believe if one takes a normal substitution-based operational semantics, because it seems that the context captured and bound to base is arbitrarily large. However, this context is mostly unevaluated code; the only heap-allocated values that are reachable from it are those that were evaluated before the evaluation of base (and used in the program after the evaluation of base). Assuming that base is evaluated early in the program, we conclude that there are few heap-allocated values reachable from its continuation. In contrast, the previous implementation of isolate could capture a context that has many heap-allocated values reachable from it (because we could evaluate isolate f late in the program and deep in a call stack), which would all remain reachable during the evaluation of f x. [We’ll return to this point later, as it is taking a slightly MLton-esque view of the evaluation of a program, and may not apply as strongly to other implementations (e.g., SML/NJ).]

Now, once we throw to base and begin executing f x, only the heap-allocated values reachable from f and x and the few heap-allocated values reachable from base are retained by the garbage collector. So, it seems that base works as an empty context.

But, what about the continuation returned from isolate f? Note that the continuation returned by isolate is one that receives an argument x and then throws to base to evaluate f x. If we used a CPS-translation implementation (and assume sufficient beta-contractions to eliminate administrative redexes), then the original continuation passed to isolate (i.e., the continuation bound to k1) will not be free in the continuation returned by isolate f. Rather, the only free variables in the continuation returned by isolate f will be base and f, so the only heap-allocated values reachable from the continuation returned by isolate f will be those values reachable from base (assumed to be few) and those values reachable from f (necessary in order to execute f at some later point).

But, MLton doesn’t use a CPS-translation implementation. Rather, at each call to callcc in the body of isolate, MLton will copy the current execution stack. Thus, k2 (the continuation returned by isolate f) will include execution stack at the time of the call to isolate f — that is, it will include the original continuation of the call to isolate f. Thus, the heap-allocated values reachable from the continuation returned by isolate f will include those values reachable from base, those values reachable from f, and those values reachable from the original continuation of the call to isolate f. So, just holding on to the continuation returned by isolate f will retain all of the heap-allocated values live at the time isolate f was called. This leaks space, since, semantically, the continuation returned by isolate f only needs the heap-allocated values reachable from f (and base).

In practice, this probably isn’t a significant issue. A common use of isolate is implement abort:

The continuation returned by isolate th is dead immediately after being thrown to — the continuation isn’t retained, so neither is the semantic garbage it would have retained.

But, it is easy enough to move onto the empty context base the capturing of the context that we want to be returned by isolate f:

This implementation now has the right space behavior; the continuation returned by isolate f will only retain the heap-allocated values reachable from f and from base. (Technically, the continuation will retain two copies of the stack that was in place at the time base was evaluated, but we are assuming that that stack small.)

One minor inefficiency of this implementation (given MLton’s implementation of continuations) is that every callcc and throw entails copying a stack (albeit, some of them are small). We can avoid this in the evaluation of base by using a reference cell, because base is evaluated at the top-level:

Now, to evaluate base, we only copy the stack once (instead of 3 times). Because we don’t have a dummy continuation around to initialize the reference cell, the reference cell holds a continuation option. To distinguish between the original evaluation of base (when we want to return the continuation) and the subsequent evaluations of base (when we want to evaluate a thunk), we capture a (unit -> unit) option continuation.

This seems to be as far as we can go without exploiting the concrete implementation of continuations in MLtonCont. Examining the implementation, we note that the type of continuations is given by

and the implementation of throw is given by

Suffice to say, a continuation is simply a function that accepts a thunk to yield the thrown value and the body of the function performs the actual throw. Using this knowledge, we can create a dummy continuation to initialize baseRef and greatly simplify the body of isolate:

Note that this implementation of isolate makes it clear that the continuation returned by isolate f only retains the heap-allocated values reachable from f and base. It also retains only one copy of the stack that was in place at the time base was evaluated. Finally, it completely avoids making any copies of the stack that is in place at the time isolate f is evaluated; indeed, isolate f is a constant-time operation.

Next, suppose we limited ourselves to capturing unit continuations with callcc. We can’t pass the thunk to be evaluated in the empty context directly, but we can use a reference cell.

Note that it is important to set thRef to NONE before evaluating the thunk, so that the garbage collector doesn’t retain all the heap-allocated values reachable from f and v during the evaluation of f (v ()). This is because thRef is still live during the evaluation of the thunk; in particular, it was allocated before the evaluation of base (and used after), and so is retained by continuation on which the thunk is evaluated.

This implementation can be easily adapted to use MLton’s primitive stack copying operations.

In essence, Thread.copyCurrent copies the current execution stack and stores it in an implicit reference cell in the runtime system, which is fetchable with Thread.savedPre. When we are ready to throw to the isolated function, Thread.copy copies the saved execution stack (because the stack is modified in place during execution, we need to retain a pristine copy in case the isolated function itself throws to other isolated functions) and Thread.switchTo abandons the current execution stack, installing the newly copied execution stack.

The actual implementation of MLton.Cont.isolate simply adds some Thread.atomicBegin and Thread.atomicEnd commands, which effectively protect the global thRef and accommodate the fact that Thread.switchTo does an implicit Thread.atomicEnd (used for leaving a signal handler thread).

It is perhaps interesting to note that the above implementation was originally derived by specializing implementations of the MLtonThread new, prepare, and switch functions as if their only use was in the following implementation of isolate:

It was pleasant to discover that it could equally well be derived starting from the callcc and throw implementation.

As a final comment, we noted that the degree to which the context of base could be considered empty (i.e., retaining few heap-allocated values) depended upon a slightly MLton-esque view. In particular, MLton does not heap allocate executable code. So, although the base context keeps a lot of unevaluated code live, such code is not heap allocated. In a system like SML/NJ, that does heap allocate executable code, one might want it to be the case that after throwing to an isolated function, the garbage collector retains only the code necessary to evaluate the function, and not any code that was necessary to evaluate the base context.