Calling convention and method resolution
Dynamic object oriented languages rely very heavily on dynamic method resolution and calling. Dynamic method resolution is the process of finding a method to call, and the calling convention defines how arguments are transferred and control changes and returns.
To start with the "easier" problem, we'd define the calling convention first, and note that the process is very very similar when done dynamically at run-time.
A quick review of existing c-like, stack based conventions will reveal what we need to consider and solve. Coming from assembler, C uses a stack pointer to push both return address and arguments. Subsequently the function may use the the same stack to push/pop local variables, and off course it usually does calls itself.
Without going into detail, there are clear problems with this. For me the biggest is that this is not object oriented. The size of argument and frame (local) sizes of the stack are not locally known and require extensive knowledge of the compiler to extract at run-time.
The approach used a constant of the time: that a method returns to where it was called. In the face of lambdas this is not true anymore. This is linked in with the lifetime of variables that arises from having code live after the calling function has returned.
To overcome some of the issues, and in general create an object oriented solution,
a linked list approach was chosen over the traditional stack based.
In rubyx the class that represents the elements of the linked list is called Message. In fact the linked list is doubly linked and holds several other necessary data, like return address, return value and others. The arguments and local variables are separated out as separate objects. These instances are fully typed and as such the runtime information about the argument is easily retrievable.
Arguments and Locals
To detail the handling of arguments and locals we'll demonstrate the approach with the arguments and not that locals are very very similar.
The Arguments are passed in what is basically a normal object. Currently derived from NamedList, but this does not add much over basic object functionality.
Argument names, as defined in the method, map to instance variable names on the object. This approach is logically equivalent to dynamically creating a class for each methods arguments, but we only create the Type. The Type being a mapping of names to indexes of where the variables are stored in the object.
To have the information available at run-time the object-type of the Arguments instance has to be changed dynamically in the method setup. As the type is known in in fact stored in the Method, this is quite easy.
When allocating the objects needed, ie Message, Arguments and Locals, we pay some
run-time price for this more explicit handling of information.
But it is not as much as one might think. All the objects are of identical size and can be kept in a linked list. "Allocation" is then just grabbing the first link from a known (compile-time known) place and relinking. About 20 risc instructions in total.
Also we need to update the type information in the objects as stated above, but this too is not much. (older) Test have shown that the calling convention is about 30-50% slower than C. This makes it very much faster than Interpretation, even taken into account the since added functionality. I would guess by now it is about 2x slower than c, which is very acceptable, given that it buys us many a simplification regarding lambdas, exceptions and fibers.
Transferring control is then quite simple, basically we jump to the next methods binary address. Before the actual jump, we have to
- Store the return address
- Swap the current Message with the next one that we prepared
This is quite equivalent to using the stack, and amounts to only one or two instructions more.
Returning from a Method is equally simple. The convention is that any possible return value will have been stored in the current Message before the ReturnSequence is initiated. Then we:
- Move the return address from the message to a register.
- Swap the previous message in (replacing the current)
- Jump to the stored address
To summarise, we
- use typed objects to store information
- use a linked list
- are well prepared for the future, lambdas etc
- can easily implement fibres
- do not have stack as a separate memory area, or concern
- have performance closer to c than mri
Which may in turn lead to an elaborate inlining scheme, thus eliminating the small performance hit completely.