Question

I want self-education purpose implement a simple virtual machine for a dynamic language, prefer in C. Something like the Lua VM, or Parrot, or Python VM, but simpler. Are there any good resources/tutorials on achieving this, apart from looking at code and design documentations of the existing VMs?

Edit: why close vote? I don’t understand – is this not programming. Please comment if there is specific problem with my question.

Answer 1

I assume you want a virtual machine rather than a mere interpreter. I think they are two points on a continuum. An interpreter works on something close to the original representation of the program. A VM works on more primitive (and self-contained) instructions. This means you need a compilation stage to translate the one to the other. I don’t know if you want to work on that first or if you even have an input syntax in mind yet.

For a dynamic language, you want somewhere that stores data (as key/value pairs) and some operations that act on it. The VM maintains the store. The program running on it is a sequence of instructions (including control flow). You need to define the set of instructions. I’d suggest a simple set to start with, like:

• basic arithmetic operations, including arithmetic comparisons, accessing the store
• basic control flow
• built-in print

You may want to use a stack-based computation approach to arithmetic, as many VMs do. There isn’t yet much dynamic in the above. To get to that we want two things: the ability to compute the names of variables at runtime (this just means string operations), and some treatment of code as data. This might be as simple as allowing function references.

Input to the VM would ideally be in bytecode. If you haven’t got a compiler yet this could be generated from a basic assembler (which could be part of the VM).

The VM itself consists of the loop:

1. Look at the bytecode instruction pointed to by the instruction pointer.
2. Execute the instruction:
   * If it's an arithmetic instruction, update the store accordingly.
   * If it's control flow, perform the test (if there is one) and set the instruction pointer.
   * If it's print, print a value from the store.
3. Advance the instruction pointer to the next instruction.
4. Repeat from 1.

Dealing with computed variable names might be tricky: an instruction needs to specify which variables the computed names are in. This could be done by allowing instructions to refer to a pool of string constants provided in the input.

An example program (in assembly and bytecode):

offset  bytecode (hex)   source
 0      01 05 0E         //      LOAD 5, .x
 3      01 03 10         // .l1: LOAD 3, .y
 6      02 0E 10 0E      //      ADD .x, .y, .x
10      03 0E            //      PRINT .x
12      04 03            //      GOTO .l1
14      78 00            //      .x: "x"
16      79 00            //      .y: "y"

The instruction codes implied are:

"LOAD x, k" (01 x k) Load single byte x as an integer into variable named by string constant at offset k.
"ADD k1, k2, k3" (02 v1 v2 v3) Add two variables named by string constants k1 and k2 and put the sum in variable named by string constant k3.
"PRINT k" (03 k) Print variable named by string constant k.
"GOTO a" (04 a) Go to offset given by byte a.

You need variants for when variables are named by other variables, etc. (and the levels of indirection get tricky to reason about). The assembler looks at the arguments like “ADD .x, .y, .x” and generates the correct bytecode for adding from string constants (and not computed variables).