blog-static/content/blog/07_compiler_runtime.md

title	date	draft	tags
Compiling a Functional Language Using C++, Part 7 - Runtime	2019-08-06T14:26:38-07:00	true	C and C++ Functional Languages Compilers

Wikipedia has the following definition for a runtime:

A [runtime] primarily implements portions of an execution model.

We know what our execution model is! We talked about it in Part 5 - it's the lazy graph reduction we've specified. Creating and manipulating graph nodes is slightly above hardware level, and all programs in our functional language will rely on such manipulation (it's how they run!). Furthermore, most G-machine instructions are also above hardware level (especially unwind!).

Push and Slide and other instructions are pretty complex. Most computers aren't stack machines. We'll have to implement our own stack, and whenever a graph-building function will want to modify the stack, it will have to call library routines for our stack implementation:

void stack_push(struct stack* s, struct node_s* n);
struct node_s* stack_slide(struct stack* s, size_t c);
/* other stack operations */

Furthermore, we observe that Unwind does a lot of the heavy lifting in our G-machine definition. After we build the graph, Unwind is what picks it apart and performs function calls. Furthermore, Unwind pushes Unwind back on the stack: once you've hit it, you're continuing to Unwind until you reach a function call. This effectively means we can implement Unwind as a loop:

while(1) {
    // Check for Unwind's first rule
    // Check for Unwind's second rule
    // ...
}

In this implementation, Unwind is in charge. We won't need to insert the Unwind operations at the end of our generated functions, and you may have noticed we've already been following this strategy in our implementation of the G-machine compilation.

We can start working on an implementation of the runtime right now, beginning with the nodes:

{{< codelines "C++" "compiler/07/runtime.h" 4 50 >}}

We have a variety of different nodes that can be on the stack, but without the magic of C++'s vtable and RTTI, we have to take care of the bookkeeping ourselves. We add an enum, node_tag, which we will use to indicate what type of node we're looking at. We also add a "base class" node_base, which contains the fields that all nodes must contain (only tag at the moment). We then add to the beginning of each node struct a member of type node_base. With this, a pointer to a node struct can be interpreted as a pointer to node_base, which is our lowest common denominator. To go back, we check the tag of node_base, and cast the pointer appropriately. This way, we mimic inheritance, in a very basic manner.

We also add an alloc_node, which allocates a region of memory big enough to be any node. We do this because we sometimes mutate nodes (replacing expressions with the results of their evaluation), changing their type. We then want to be able to change a node without reallocating memory. Since the biggest node we have is node_app, that's the one we choose.

Finally, to make it easier to create nodes from our generated code, we add helper functions like alloc_num, which allocate a given node type, and set its tag and member fields appropriately. We don't include such a function for node_data, since this node will be created only in one possible way.

Here's the implementation: {{< codelines "C" "compiler/07/runtime.c" 6 40 >}}

We now move on to implement some stack operations. Let's list them:

• stack_init and stack_free - one allocates memory for the stack, the other releases it.
• stack_push, stack_pop and stack_peek - the classic stack operations. We have _peek to take an offset, so we can peek relative to the top of the stack.
• stack_popn - pop off some number of nodes instead of one.
• stack_slide - the slide we specified in the semantics. Keeps the top, deletes the next several nodes.
• stack_update - turns the node at the offset into an indirection to the result, which we will use for lazy evaluation (modifying expressions with their reduced forms).
• stack_alloc - allocate indirection nodes on the stack. We will use this later.
• stack_pack and stack_split - Wrap and unwrap constructors on the stack.

We declare these in a header: {{< codelines "C" "compiler/07/runtime.h" 52 67 >}}

And implement them as follows: {{< codelines "C" "compiler/07/runtime.c" 42 116 >}}

Let's now talk about how this will connect to the code we generate. To get a quick example, consider the node_global struct that we have declared above. It has a member function, which is a function pointer to a function that takes a stack and returns void.

When we finally generate machine code for each of the functions we have in our program, it will be made up of sequences of G-machine operations expressed using assembly instructions. These instructions will still have to manipulate the G-machine stack (they still represent G-machine operations!), and thus, the resulting assembly subroutine will take as parameter a stack. It will then construct the function's graph on that stack, as we've already seen. Thus, we express a compiled top-level function as a subroutine that takes a stack, and returns void. A global node holds in it the pointer to the function that it will call.

When our program will start, it will assume that there exists a top-level function f_main that takes 0 parameters. It will take that function, call it to produce the initial graph, and then let the unwind loop take care of the evaluation.

Thus, our program will initially look like this: {{< codelines "C" "compiler/07/runtime.c" 154 159 >}}

As we said, we expect an externally-declared subroutine f_main. We construct a global node for f_main with arity 0, and then start the execution using a function eval. What's eval, though? It's the function that will take care of creating a new stack, and evaluating the node that is passed to it using our unwind loop. eval itself is pretty terse:

{{< codelines "C" "compiler/07/runtime.c" 144 152 >}}

We create a fresh program stack, start it off with whatever node we want to evaluate, and have unwind take care of the rest.

unwind is a direct implementation of the rules from Part 5:

{{< codelines "C" "compiler/07/runtime.c" 118 142 >}}

We can now come up with some simple programs. Let's try writing out, by hand, main = { 320 + 6 }. We end up with:

{{< codeblock "C" "compiler/07/examples/runtime1.c" >}}

If we add to the bottom of our main the following code:

printf("%d\n", ((struct node_num*) result)->value);

And compile and run our code:

gcc runtime.c examples/runtime1.c
./a.out

We get the output 326, which is exactly correct!

We now have a common set of functions and declarations that serve to support the code we generate from our compiler. Although this time, we wrote out f_main by hand, we will soon use LLVM to generate code for f_main and more. Once we get that going, we be able to compile our code!

Next time, we will start work on converting our G-machine instructions into machine code. We will set up LLVM and get our very first fully functional compiled programs in Part 8 - LLVM.