Abstract Machines: Interpreters for Computer Scientists
Nov 09, 2020 · 21 min

# Introduction

So im a PhD student now, so I need to write about cool computer science things! As part of my learnings, I have been writing a bunch of 'Abstract Machines'. I think of them as how computer scientists do programming languages research without computers. I mean, they have to use something to write their super complex papers.

"Computer Science is no more about computers than astronomy is about telescopes"

• Edsger Dijkstra

We love computers, but they are merely a tool of computing. The real study of computation can be done without them, and the theories of how programming languages function is not excluded from that. So how do computer scientists study interpreters without a computer? The theory of abstract machines is one of the more popular ways.

# How to Compute

What does it mean to compute something? Humans are pretty good at just looking at things and formulating an answer. For example, a person does not use a sorting algorithm when matching socks after their laundry is finished. But computers can't just intuit a solution. They are given precise instructions on what to do to accomplish something. So how can we model that, and use it to inform the science behind computation?

'Abstract Machines' were created to model real computation strategies. These are functions that take program states, and return some value. Program states can be composed of many different things. The simplest abstract machines simply use the current point that the program is at. Others include a mapping of variables to values, so we can keep state around. We will describe such machines, and what kind of languages they can describe.

These machines, in practice, are interpreters. They are called 'abstract' because the theory on them is not specific to any exact language. You can make an abstract machine for whatever you could want: Lisp, Java bytecode, RISC-V assembly language...

# Digression: Operational Semantics

Operational Semantics are how we can write semantics of a language using pen and paper. Using a simple syntax, we can get the idea of 'if the syntax looks like this, it can be evaluated into this'.

The idea of semantics is separate from execution of a program. Semantics describe what something 'means', based on how it looks. Here are two different kinds of 'operational semantics', big-step and small-step.

## Big Step

Big step evaluations have the type State -> Value, meaning that you give them a state, and it will tell you which value it exactly evaluates to. These are nice and simple generally, because they give you the values in a single step.

To use a small example expression, if cond then e_true else e_false there are two rules that will be used.

1. If cond evaluates to true, then the result is what e_true evaluates to.
2. If cond evaluates to false, then the result is what e_false evaluates to.

These rules assume that we have some way to fully evaluate whatever cond is. But whatever that way is is unimportant to the rule of evaluating if statements. This separation is very important for operational semantics. We can write small rules for specific parts of the language, all of which get composed together into a fully formalized machine.

I like using big step when thinking exactly about what expressions do.

## Small Step

Small step semantics have the type State -> State. They will show you what to do step by step to evaluate a term. They work iteratively, explicitly, to show how a given state is computed.

For example, if the term is if cond then e_true else e_false, There will be a few different rules.

1. If cond is an atomic value, and that value is true, the resulting state is e_true.
2. If cond is an atomic value, and that value is false, the resulting state is e_false.
3. If cond is not an atomic value, then evaluate it to cond' , and return if cond' then e_true else e_false.

An atomic value is a value that can not be broken up any more. This means a datatype, not a complex expression. In these simple machines, the only atomic datatype is a number. Don't tell computer scientists about quarks, they may go insane.

By going from a state to a next state, small step mechanics more closely follow how our computers work. They dont evaluate to values directly, they just... keep going.

This may confuse a new reader, if it keeps going, how do we know its done? Big step rules directly result in a value, full stop. How do we know a state in small step semantics is the one with the value? We use what is called a 'fixpoint', or more simply, we evaluate until there isnt a meaningful change in the state after running. If we evaluate a math expression enough times, it will decompose into a single number, upon which evaluation will lead to itself. That means there is no more work to be done, and evaluation stops.

The same is true in a language like C, after the final instruction in main, if we try to evaluate any more, nothing will happen, theres nothing left to do. That is a fixpoint.

Small step semantics are a bit more precise in my opinion, and they are much easier to translate to real interpreters. But it is very useful to understand both styles, they have different uses. Using small steps, we can also more closely trace how an expression is evaluated.

# A Simple Abstract Machine, C

I dont mean the the C language, especially because I would not use simple to describe it! C in this case stands for 'control'. You may remember things like 'if' are called 'control flow operators'. Control is the currently running 'thing' in your program. The if operator, is a way to change the control based on a condition. This machine will be called the C machine because you only need control to represent the state of the entire program.

C machines are not capable of much, only simple rewriting of expresions, because they dont have any information other than the program's control itself to go off of. One kind of language we can formalize with a C machine is that of mathematical expressions.

Here is a big-step semantics of type $MathExp \Downarrow Number$ (takes a math expression and returns a number). This means that the control we choose to use is a math expression. The result of evaluating a math expression is a number, of course, so thats the value.

$Addition:$

This rule shows how to add expressions to end up with a number. You can understand it by reading the half under the bar as 'this is what we start and end with' and the half above the bar as 'these must be true to use these semantics'. You read $\Downarrow$ as 'evalutes to'.

You can read this rule like so:

1. If we have some control that looks like $e_1 + e_2$ (this is from the bottom left, before the arrow)
2. if $e_1$ evaluates to some number $n_1$ (the first expression above the bar),
3. if $e_2$ evaluates to some number $n_2$ (the second expression above the bar),
4. if $n_1$ added to $n_2$ is equal to some number $n$ (the third expression above the bar).
5. THEN we know that the expression in step 1 evaluates to $n$.

This may seem a bit backwards, we implement adding by adding? Well, the key is that expressions are complex, they can be composed of other expressions or just values. These rules show how to do math on expressions by first evaluating them to values. Then once they are values, it is quite easy to do math operations on them.

Note the distinction between the arrow $\Downarrow$ and $=$ here. $\Downarrow$ is saying "left evaluates to right by virtue of applying this machine's rules." and $=$ is saying "you can substitute left for right".

Authors of semantics like these love to use different looking arrow symbols, they all mean the same thing. Usually in big step they use a cool down-facing arrow like $\Downarrow$. In small step they will use a more boring arrow like $\rightarrow$.

### Example

Lets evaluate a simple math expression to show how you can use these rules to prove that expressions are evaluated correctly using a machines rules.

Here, at the first, bottom most level, $e_1 = 7 + 3$ and $e_2 = 4$. Then we need to prove that 7 + 3 = 10, and we do that with another application of the addition rule! We could have chosen 3 + 4 to be be $e_2$, but it doesnt matter for addition, and we leave issues like that to a parser. I noted each level with the rule that was used to evaluate it. It is generally showed like this, but I usually dont show them. I only have so much horizontal space on this webpage!

To evaluate simple mathematical expressions, we only need a control for the state. C machines are only capable of evaluating simple programs. What if we want to add a simple programming construct like variables? For that, we need a place to store them. And so, the CE machine is born!

# Environments with the CE machine.

To evaluate variables, we need to be able to keep track of what value they hold at a given program point. in a C machine, if we are given a + 2, we have no way of know what a is, because its not a number, and we can only deal with syntax as we see it. But if we gave a machine both a + 2, the control, and a mapping {a : 4}, an environment, we can evaluate the expression to 6!

So, a big-step CE machine doesnt just have MathExp (C) for state anymore, we need an Env (E) to accompany it! The function is now of type $(MathExp \times Env) \Downarrow Number$. This means that our machine will take 2 arguments, a math expression and an environment, and it will return a computed number.

$Var-Lookup:$

This simply states that if the expression is a variable, not an artithmetic expression, the value it returns is what the environment says it is. We use the greek letter rho ('ρ') (not the letter 'p') to represent the environment. This is what is used in the literature, and its always best to follow along with norms to avoid confusion! For the history on why they used $\rho$ for this task, you will have to ask someone smarter than me.

We dont have variable binding yet, but you can imagine an expression like: m*c*c with an environment {m : 12 c : 299792458}, which will return some number for us. We have made a calculator with predefined constants! We can put pi in there, or tau if you are a lunatic...

# Real Looking Programs

Wow, it seems like we can do a lot with just a control and an environment. We could further extend this to things like variable assignment (again, these are big-step semantics):

$\mathbf{Assignment:}$

This tells us that when we assign a variable, it will be available in the next expression.

It seems that these two components of a state are enough to be wildly dangerous with. Soon enough, your semantics will be modeling a real language, with conditions, and inequalities, oh my! But, there are two big problems that you may face.

First, your big step semantics may limit you. Their stringent requirements that their sub-expressions be terminating can cause issues in a turing-complete environment. And a small-step semantics can provide more granularity of control. For that reason, we will be using small-step semantics from now on.

Also, as your language gets more and more complicated, the constructs of control and environment may be overburdened with responsibility. What can we do to simplify our understanding of these machines?

# What next? Kontinuations! CEK

One way to alleviate the burden is to add a new construct to our machines: The continuation. A continuation is a way of showing 'future work' to be done. For example, when we implement our assignment operation, we can delegate the work of executing the second expression to the continuation:

$\footnotesize{\mathbf{Assign-Exp}}:$ $\footnotesize{\mathbf{Assign-Kont}}:$

There is quite a bit more to unpack here!

First, we switched to a small step semantics. Instead of returning a fully evaluated value as before, we now return states. Big step semantics are nice for simple programs, but when things get more complex, we want finer-grained control, so we switch to small-step.

Next, our state has a third argument. Our type signature is now $(C, E, K) \rightarrow (C, E, K)$. The third argument is a continuation, it is useful in small step semantics, so we can focus on evaluating expressions. When we are done, we know what to do next.

Finaly, The above semantics show that when we see an assignment expression, we simply return the $e_1$ expression as the control for the next state. The key is that we save a new 'continuation frame' along with that control. This signifies that after that expression is fully evaluated, we continue by assigning it to a variable, and then execute the expression $e_2$.

Another notable thing is that, inside of that 'assign' continuation frame, we store the old continuation. This recursive format is necessary when nesting complex expressions. Let's use an example with nested assignment expressions, with the recursive nature of the continuations being our life-jacket.

## Example

x := 4 in (x + 1), we would start with the assign-exp rule, which simply gives us a new state, it doesnt directly evaluate anything to a base value. The new state says that we need to evaluate 4 with the same environment as before, but we change the continuation, to store the values in the we need to use after the 4 is fully evaluated (which it is, but we cant see that far ahead! One small step at a time!). After the assign-exp a state that looks like (4 , env , assign(...)) is returned. We immediately see that the assign-kont rule can be applied, as the control is a number, and we have an assign continuation in the K position. So by applying that, a state (x + 1, env2 , old_kont) is given to us, and with x in the environment mapped to 4.

What is old_kont in this example? As you see from the assign-exp rule, one of the main things that a continuation does is store the previous one, like a linked-list. Continuations are the 'rest of the program', so after the current continuation is used, we do what the old continuation was.

For example, an program that looks like:

Here is what the states will look like over the course of 3 applications of assign-exp:

(x := (y := (z := 3 in z) in (y + y)) in x , {} , NONE)
-> ((y := (z := 3 in z) in (y + y)) , {} , assign(x , x , NONE))
-> ((z := 3 in z) , {} , assign(y , y + y , assign(x , x , NONE)))
-> (3 , {} , assign(z , z , assign(y , y + y , assign(x , x , NONE))))


Then if we apply assign-kont a couple times, with omitted addition and var-lookup rules:

-> (z , {z : 3} , assign(y , y + y , assign(x , x , NONE)))
-> (3 , {z : 3} , assign(y , y + y , assign(x , x , NONE)))
-> (y + y , {z : 3 , y : 3} , assign(x , x , NONE))
...
-> (6 , {z : 3 , y : 3} , assign(x , x , NONE))
-> (x , {z : 3 , y : 3 , x : 6} , NONE)
-> (6 , {z : 3 , y : 3 , x : 6} , NONE)


### Interlude: Injection

Note above that there is an 'initial' state that we use when we start evaluating an expression. We wrap the expression with an empty environment ('{}'), and an 'empty' contination ('NONE'). This is usually done with whats called an injection function:

This gives us the initial state from an expression. The initial state can of course be changed to your needs. adding initial environment values or constants, for example.

### Back To The Example

We started with a NONE continuation, meaning have nothing to do next (end of program!). We then slowly build up continuations. Continuations are turned to for guidance once control is atomic. Atomic here means that it cant be broken down any more (computer scientists dont have quarks yet). Once we evaluate a continuation, the cycle continues, evaluate the control until it is atomic again. Eventually (if your program terminates!), we will be faced with an atomic control, and a NONE continuation. This means that we have finished evaluation, as there is nothing left to do! That is known as a fixpoint, which is a term taken from math meaning, if we run it another step, nothing will change: f(x) evaluated to x.

This is a good example of the differences between small and big step evalaution models. Small step is a scalpel, where big is a cudgel. They may be more effective than the other, depending on what youre doing.

# Other Continuation Types

We have demonstrated two different continuation types, "NONE", and "ASSIGN". One signifies the end of computation, and one tells us how to add variables to the environment. What if we want to add a conditional to our language? It is very easy with our language!

Lets compute this text:

This will evaluate c, and once a value is computed, it checks its value. If the value is 0, it will take the else branch, otherwise it will take the then branch. If we had a more complex language with booleans, we could use those. Or really, we could do almost anything! Make your own language, make it weird!

$\text{if-exp}:$ $\text{if-kont-t}:$ $\text{if-kont-f}:$

Here, when it is time to check the continuation for guidance, we will next evaluate one of the branches from the ifk continuation 'frame' based on the controls value.

The idea of small step + continuation can be very similar to big-step. this CEK machine is basically saying 'evaluate this to a value, then do something else'. Big-step says that but in a slightly more compact way, involving more inference rules on the top half of the big bar.

## Another Example

Lets do another simple example of continuations, just to hammer them in!

We will show that:

(x := (if (1 + 1) then 7 else 12) , {} , NONE)
-> (if (1 + 1) then 7 else 12 , {} , assign(x, x, NONE))
-> ((1 + 1) , {} , ifk(7, 12, assign(x, x, NONE)))
-> (2 , {} , ifk(7, 12, assign(x, x, NONE)))
-> (7 , {} , assign(x, x, NONE))
-> (x , {x : 7} , NONE)
-> (7 , {} , NONE)


### Fun Insight!

If you have been paying close attention to the continuations, they kind of act like a fine-grained stack. Traditional stacks are at the level of funtions. You call a function, and a stack-frame is pushed. After a return is issued (in our case, reaching an atomic value), that frame is popped. The idea with continuations here is very similar, but at the level of instructions (like conditionals, assignments, etc.), not just functions.

I didnt want to get into the weeds of real progrmaming in this post, but if that helps you understand more, great! If not, dont fret! But think on the recursive nature of continuations, it can really give shed new light on how to model computation!

In practice, recursive continnuations can be made efficient by using tail-call eliminations, but thats a subject for a whole 'nother post!

# The Big Idea

The abstract machines are the C, CE, and CEK ideas that we have talked about. They are the types of state used when implementing an abstract machine. The resulting machine, however, is a concrete interpreter for your programming language.

So what is wih all this trouble? Creating a tiny interpreter is easy! Using abstract machines offers us a way to build new features incrementally, and concisely.

These Abstract Machines provide a way to form an independent interpreter for your language, that doesnt necessarily depend on the runtime of your host language. Also, using operational semantics can help you to formalize how your language evalautes, and be sure that your implementation is correct!

But more than all of that, this is how computer scientists can model real languages. To accomodate that, other components can be added to the state, such as a 'store', for modeling the heap. A CESK machine changes the environment to be a mapping from variables to 'addresses', and the store becomes a mapping from addresses to the values. This is helpful for modeling mutation, and has implications in computability when doing static program analysis.

# The Future

Adding even more complex language features is a necessity for formalizing a real language. One key feature is of exception handling: try-catch and the like. A solution used for advanced control flow such as try-catch is to bring continuations to first-class status. Continuations are very powerful constructs, and can be used to implement exception handling schemes. I am writing another post on what continuations are, becuase they are very poorly known in the world outside of programming panguages research.

Abstract Machines can themselves be abstracted, to form interpreters that are non-deterministic. These interpreters can be used to see what possible states a program can reach. Using this information, static analyses of programs can be conducted, meaning optimizations, warnings, etc.. The ideas of abstract AMs is still new, and a topic of current research.

# Conclusion

Abstract Machines are a way to formally define computation for terms. When mapped to 'the real world' that means an interpreter for your programming language. They are a wonderful way of explicitly showing how programs are executed, and computation is performed.

# Appendix

 $\rho:$ 'rho', is an environment, used to map variables to values. $\kappa:$ 'kappa', is a continuation, which represents the rest of the program. We consult this after the current control is fully evaluated. $\Sigma:$ 'Sigma' (uppercase). Although not used in this post, this character often is used to represent the State type of the machine. $\varsigma:$ 'sigma' (lowercase in 'final word position'). This is known as 'varsigma'. You may find this used as an instance of a state. A small step machine goes from some state to a next state, so we can codify that as $\sigma:$ 'sigma' (lowercase). Although not used in this post, this character is used to represent the store component of a machine. If we would have made a CESK machine, then the S would be this.