On Designing Interfaces

I talked a little bit about ATs in an earlier post. To summarize, ATs serve three different roles in the original Lush interpreter, namely

1. As generic, unique identifiers for Lisp objects
   
2. As carriers of runtime type information
3. As carriers of information for memory management (reference counts)
   

In the branch that I am currently working on, the third point is no longer true and ATs serve only the first two purposes. In this posting I want to talk about the first purpose.

There are many type-generic functions in the interpreter that may work with many or all kinds of objects (creating a string representation, writing an object to a file, computing a hash code, etc.). Many other functions work with objects of particulars type only (open a file given a name (string), slicing an array, arithmetic functions, etc.). And a third kind of function are found in the class structure of Lisp objects. They are type specific versions of functions that should work with all objects (dispose, creating a string representation, listeval, etc.).

The natural signature of type-generic functions include an AT argument pointing to the object in question. They may be implemented as dispatchers to type-specific functions of the third kind. Function signatures for the second kind may or may not include an AT.

When you look at the original Lush interpreter implementation you will find that functions of the second and third kind do have generic signatures as well. The ones of the second kind typically do a runtime type-check, the ones of the third do not. I consider this bad design. If a function works with arguments of a particular type only, then the signature should say so and the implementation should not waste (run)time doing type checks. This is one kind of changes that I have been making since I started hacking Lush, and there's still a lot of work to be done on overhauling internal interfaces.

Yesterday I was churning through the dispose functions, which are functions of the third kind. Objects with non-trivial dispose functions need a finalizer to invoke the dispose function. Lush also has a function delete to explicitly destroy a Lisp object, no matter if it is still referenced or not. Delete calls an object's dispose function and then modifies the AT to make it look like a NIL. So a dispose needs to be idempotent since it may be called twice (once by delete and a second time by the finalizer).

Since I was going to look at all dispose functions I decided to change their signatures from generic to type specific along the way. When I got to dispose for OO objects, I hit a snag: This dispose invokes a user-defined destructor (if one has been defined), and at that point, I need to know the object's AT because the destructor is a Lisp function and evaluate has a generic signature. After making dispose type-specific, however, I don't have the object's AT anymore.

That led me to wonder if maybe I was wrong and there is some value in having all signatures generic. Anyway, I fixed it by adding a "back pointer" to the OO object structure (a pointer that points back to the AT referencing the object), which is not nice. Perhaps the dispose interface still isn't quite right?

Crash course in Debugging

Debugging sucks. After implanting the new memory manager into the interpreter nothing worked at first. And after over a week spent on debugging, it is still not working. But yesterday I had a good run and slayed over half a dozen bugs. Feels good.

This is my crash course in GDB. I used GDB before occasionally, but I was always starting from zero because I use it so infrequently. This time around there are so many bugs to hunt down that I really need to use a lot of GDB's capabilities and I need to use it for a while! That's how you learn a tool really well.

Easier tracking of transient objects

My API for tracking transient objects in MM is a little different from Nickle's. Consider the three following functions.

char *a (int n)
{
assert(n>=0);
char *s = mm_allocv(mt_blob, n+1);
s[n] = '\0';
return s;
}

char *b (date_t d)
{
char *s = a(127);
sprintf(s, "Hello world on %s.\n", date2str(d));
return s;
}

char *c (date_t d)
{
MM_ANCHORED;
char *s = a(127);
sprintf(s, "Hello world on %s.\n", date2str(d));
MM_RETURN(s);
}

Functions b and c are almost the same, except that in c the body is what I call an "anchored scope". Neither b nor c call an allocation function directly, but a does.

Now, will mm_allocv in a push the address of the allocated memory onto the transient object stack or not? It depends. If you are not in an anchored scope and call a or b, then it won't. If you call c, it will. In other words, the allocation functions have context-dependent behavior: they push new addresses when executed in an anchored scope and don't push otherwise. Along the same lines, when exiting an anchored scope with MM_RETURN, the returned object will be pushed only if the return is into an(other) anchored scope.

I believe this behavior is more useful than that of the Nickle system. There are lots of little functions in Lush who create a new object. They typically do one or a small number of allocations and then initialize the object. I do not need to fortify all these functions with MM_ANCHORED, it is good enough to do it to the functions that call such constructor functions. In essence, I believe I will need to use MM_ANCHORED and company less often and will find it easier to write correct code when using it.

Destructors and Finalizers

As Hans Boehm argues in a recent paper, finalizers should run asynchronously to the client code. I will support this in my memory manager by queuing up finalization-ready objects and leave it to the client to actually run the finalizers through the API mm_run_pending_finalizer.

Lush "OO objects" may have a destructor method. For stack-allocated objects, the destructor method should be invoked when the object goes out of scope. For heap-allocated objects, the destructor method should invoked after the object becomes unreachable. In other words, "OO objects" need a finalizer, and one of its jobs is to call the object's destructor method. This makes clear that, when there are several finalization-ready objects, the order of finalizer invocation is important.

Current Lush has a facility for registering "finalizers" with individual objects. These are used to notify other objects that hold a weak reference to an object. For instance, a hash table may hold weak references to key objects, and events in the event queue hold a weak reference to an event handler object. The "finalizers" registered by the hash table and by the event queue do not look at the object, but only need to know an object's identity to update the hash table, or event queue, respectively. I would call them "notifiers" rather than "finalizers".

To support this mechanism for maintaining weak references I am extending the memory manager API by a new function mm_notify. This function will just mark a managed object. When the object has become unreachable and when it is about to be reclaimed by the memory manager, a notification function is called with the object's address as argument. The notification function is not specific to an individual object or an object's type and it needs to be given to mm_init when the system is initialized.

Would you like some cream with that?

Just a quick update: The memory manager now organizes small objects of same memory type in blocks; a client needs to register a memory type before first allocation; for larger objects the memory type is stored right in front of the object itself (similar to Nickle's scheme). This allows me to just store the addresses instead of pairs of object and memory type addresses. Here is a brief documentation of the current API.

The next steps are (1) extension of the API by macros for tracking transient objects, (2) tearing open and stitching back together the Lush interpeter, (3) debugging, debugging, debugging, (4) implementing asynchronous collect.

Since there are no special functions for dealing with ATs anymore, I am feeling this could become a library...