Thinking > story
Prototyping an interpreter for a small language using Makam

In this post, we'll implement a toy interpreter for a small functional language. We will use Makam, a metalanguage that helps in the 'initial spiking phase' of designing a new language, allowing for a tight feedback loop and for iterating quickly.

tests : testsuite. %testsuite tests.

Say you are designing and implementing a new language: there is a class of programs that are hard to write in languages that are presently available and you want better ways to express them.

There are a lot of decisions to make. For starters, what should the constructs of the language be? How do these constructs enable writing the example programs that you have in mind? Which constructs should be the "core" ones of the language, and which ones should be defined in terms of them?

What does it mean to use the language? How do you write programs in it — what's the syntax like, what information can one get about their programs (and how much of it can be inferred)? What do the constructs of the language mean — how do you compute with them, how do they relate to the constructs in existing languages?

Coming up with answers to these questions is an iterative process: you can start with some answers, try to write example programs, see what works and what does not, and adapt accordingly. Implementing the language is quite crucial to this process: actually using the language reveals patterns that are important but that you couldn't necessarily have found otherwise — so that informs how to refine the language further and what constructs to add.

Still, implementing a language takes a long time, which hinders this experimentation and refinement process. There is typically a long feedback loop involved between having a new language design idea and having a working (even toy-ish) implementation of it... but there doesn't have to be!

Minimizing this feedback loop is exactly why I have been working on the design and implementation of Makam, a metalanguage that helps with the design and implementation of (prototypes of) new languages. This way it allows you to iterate, validate and refine your language ideas quickly.

Makam is a dialect of lambda-Prolog and is hence a higher-order logic programming language (more on what that means later). This post contains Makam code that you can run using the Play button on the bottom-right corner; you can also edit the final code block to try out your own examples (the second button will take you there). You can also follow the installation instructions for Makam and run this post on your machine if you download the source code for this post and do "makam makam-tutorial-01.md -".

In this series of posts, we will use Makam to prototype various parts of a toy programming language. We will also talk through the current set of answers of the lambda-Prolog/Makam language design, in terms of what the base constructs are, and what can be programmed using those. In this first part, we will focus on writing an interpreter for a very small language, and only talk about the (non-lambda) Prolog part of Makam. More to come!

Expressing the main constructs of our language: Abstract syntax

Let's start defining and implementing our toy language. First of all, we will need to decide what our base constructs are. To keep things simple let's start with this:

  • String constants like "foo", "bar"
  • Integer constants like 5, 42
  • Boolean constants, namely true and false
  • An expression to add two integers or two strings together
  • Array literal expressions, like [40 + 2, "foo"]
  • Record literal expressions, like { foo: "bar" }

Here is a more formal way where you might see this kind of definition of language constructs on paper:

$$\begin{array}{llll} e & \text{(expressions)} & ::= & s \; | \; n \; | \; b \; | \; e_1 + e_2 \; | \; [ e_1, \cdots, e_n ] \; | \; \\ & & & \{ s_1: e_1, \cdots, s_n: e_n \} \\ s & \text{(string constants)} & ::= & \cdots \\ n & \text{(integer constants)} & ::= & \cdots \\ b & \text{(boolean constants)} & ::= & \text{true} \; | \; \text{false} \end{array}$$

In this notation, expressions, string constants, integer constants etc. are different sorts — the different "kinds of things" that might be involved in the terms of our language. For example, if we were encoding an imperative language that included statements and statement blocks, we would have separate sorts for them, like:

$$\begin{array}{llll} st & \text{(statements)} & ::= & x = e; \; | \; \text{return} \; e; \; | \; \text{if} \; (e) \; \text{then} \; b_1 \; \text{else} \; b_2; \; | \; \\ & & & \text{for} \; (\text{const} \; x \; \text{in} \; e) \; b; \; | \; \text{for} \; (\text{const} \; x \; \text{of} \; e) \; b; \; | \; \cdots \\ b & \text{(blocks)} & ::= & \{ \; st_1 \; \cdots \; st_n \; \} \end{array}$$

Each alternative given for a sort is a constructor. When we read a constructor definition like $e_1 + e_2$ it means that it's a constructor for expressions that is formed by two expressions. Similarly, $\text{if} \; (e) \; \text{then} \; b_1 \; \text{else} \; b_2;$ is formed by an expression and two blocks. A constructor of the form $[ e_1, \cdots, e_n ]$ means that it is formed through a list of expressions. So the letters we give to sorts (like $e$, $s$) are a handy pun that allows us to specify the constituent parts of each constructor concisely.

Now this notation mixes a couple of things together: we are defining what the constructors are together with what is the real syntax that we will use to write down those constructors. However, we can separate those two aspects of the definition out. In terms of what the language is, the important part is what the constructors are. The syntax that we use for them is secondary: it is important in terms of actually writing down terms of the language in a way that is human-readable, but we could have different syntaxes for the same exact language.

Instead, we can separate those two concerns into two parts: one where we just give an explicit name to each constructor and describe what its constituents are (how many are there and of what sorts) — on paper, we could denote that with something like $\text{add}(e_1, e_2)$; and one where we describe what the real syntactic form for the constructor is when we write out a program in the language. When we talk about abstract syntax, we refer to the first part; and concrete syntax is the latter one.

Let's now see how we would encode these in Makam.

First of all, we define the sorts that we need, which are referred to as types in Makam. For our simple language, we just need expressions:

expr : type.

There are built-in sorts for strings and integers in Makam; booleans and lists are already defined in its standard library. Like in most functional languages, all elements of a list are of the same type. So lists of expressions are a different type than, say, lists of strings: list expr vs. list string. There's two ways to write down a list; either in the form [1, 2, 3], or in the form 1 :: 2 :: 3 :: Nil, similar to other functional languages.

With these in mind, we can define the constructors for expressions as follows. This corresponds to the definition of the abstract syntax of our language, as mentioned above:

stringconst : (S: string) -> expr.
intconst : (I: int) -> expr.
boolconst : (B: bool) -> expr.
add : (E1: expr) (E2: expr) -> expr.
array : (ES: list expr) -> expr.

So we first give the name of the constructor, like add, the arguments it takes (that is, its constituent parts), like (E1: expr) (E2: expr), and the resulting sort that it belongs to, like expr, following the arrow. The names of the arguments, like E1 and E2, are only given as documentation. This helps sometimes to disambiguate between what each different argument is — for example, we could define the if-then-else statement as:

ifthenelse :
  (Condition: expr) (Then: block) (Else: block) -> statement.

Terms built out from constructors like these correspond exactly to abstract syntax trees. For example, the abstract syntax tree for the concrete syntax 5 + 3 would be:

Abstract syntax tree

We would write this as add (intconst 5) (intconst 3) in Makam.

You might find something about the above definitions weird at first, coming from a language like Haskell from ML. Typically in functional languages we define datatypes and give all of their constructors as part of a single declaration. Here, however, we have defined new types and new constructors for those types as separate statements. In Makam, this different style of definitions allows us to define new constructors for an existing type at any point. This is a key departure of Makam/lambda-Prolog from traditional functional languages; the next post will be mostly about this feature and what it allows us to do. For the time being though, we can say that one case where this feature is useful is developing a language in stages: for example, we can define a 'base version' of a language first, and then add some extra constructs later, in a separate place, without having to change the base definition.

We have left the constructor for records out. We can view a record as a list of fields, where each field pairs together a key with a value:

field : type.
record : (Fields: list field) -> expr.
mkfield : (Key: string) (Val: expr) -> field.

And that covers all the constructors we'll define for the time being. Now let's see how to actually define computations over these terms. Our example will be an interpreter for our language that computes the value that an expression evaluates to.

Computation in logic programming

We have to pause working on our toy language implementation for a bit to first explain a little bit about how computation in Makam works.

Say that instead of using Makam, we were using a functional language. One of the main operations of functional languages is pattern-matching: we try to match a term against a pattern; if the match is successful, we proceed to take the corresponding branch. Patterns are kind of like "templates" for terms: some parts are explicitly specified, while others are allowed to be arbitrary. Another way to say this, is that if terms are like trees, patterns are like "trees with holes":

Pattern

We give names to the holes, so as to be able to refer to them — these are the pattern variables. Pattern matching basically tries to find a way to fill in these holes in the pattern so that it matches the term exactly. So its result when it's successful is an instantiation (or substitution) for the pattern variables:

Pattern

Here's an example of a query that performs pattern matching between a pattern and a term in Makam. We will talk about what queries are later on, but if you run this post right now using the Play button on the bottom-right corner, you will see that an instantiation for the pattern variables N, X is found:

pattern_match
  (add (intconst N) X)
  (add (intconst 5) (intconst 3)) ?
>> Yes:
>> N := 5,
>> X := intconst 3.

Logic programming instead allows terms to include unknown parts in them and treats unification as one of the key operations. This is the symmetric, more general, version of pattern matching: instead of having a "pattern" with potentially unknown parts on the left, and a fully known "term" on the right, we have two terms with potentially unknown parts in them, and we are trying to reconcile them against each other. This process might force instantiations on either one of them, making previously unknown parts known, or even on both of them (in different parts of them). Some things might even remain unknown after the unification. To be able to refer to them, we give names to the unknown parts — so an unknown part is a special kind of a variable, referred to as a unification variables.

Pattern

In Makam, unification variables are denoted with identifiers starting with uppercase letters, whereas the identifiers of normal term constructors start with lowercase letters. Here's an example of a query that performs unification between two terms, corresponding to the example above:

unify (add (intconst N1) X2) (add X1 (intconst N2)) ?
>> Yes:
>> X1 := intconst N1,
>> X2 := intconst N2,
>> N1 := N1,
>> N2 := N2.

(Note the color-coding on the side of codeblocks of this post: blue blocks are things that will be sent to Makam, which become green after a successful run, and grey ones are skipped. Any results from the Makam interpreter, or any errors, show up as annotations in each codeblock.)

This choice has a wide-ranging implication on how computation in logic programming actually looks. In a functional language, at the point where a function is applied, its inputs are fully known (or at least fully knowable, in a call-by-need language), whereas outputs are fully unknown, to be determined through evaluation of the function. In a logic programming language, there is no need to explicitly separate inputs from outputs: both of them could only be partially known at the point where a "function" is applied, and unification will reconcile the known and unknown parts. So instead of functions we talk about predicates: these describe relations between terms, without explicitly designating some of them as inputs and some as outputs. What is an input and what is an output depends on the arguments that the predicates are called with. Here is an example of the append predicate for lists:

append [1,2,3] [4,5,6] ZS ?
append [1,2,3] YS [1,2,3,4,5,6] ?
append XS [4,5,6] [1,2,3,4,5,6] ?

So the append XS YS ZS predicate takes three lists as arguments; the first two, XS, YS, are the operands, and ZS, the last one, is the result of the append. However, the predicate can be used not only to find the result of appending a fully-known YS to XS, but also to discover the value of XS or YS given the other operand and the result. Here is the type of append:

append : (XS: list int) (YS: list int) (ZS: list int) -> prop.

The name of the type prop comes from proposition: these are the statements that we can query upon, and might be viewed as the logic programming equivalent of the expressions of a functional programming language. So a fully applied predicate like append XS YS ZS is a proposition, and by querying about it as we did above, we are asking the Makam interpreter to find an instantiation for the unknown unification variables that makes the proposition hold.

The append queries above might be surprising at first, so let's see how append is implemented. In logic programming, we implement a predicate by defining its rules: basically, we define the cases for which a certain proposition, like append XS YS ZS holds. Each rule has a goal and optional premises, written roughly as goal :- premises (note the ":-" which can be read as "when"). The way these rules are executed is like this: given the current query Q that we are trying to solve, we attempt to unify it with the goal of each rule; if unification is successful, we proceed to the premises, treating them as subsequent queries that need to be satisfied.

The rules that make up append are these:

append [] YS YS.
append (X :: XS) YS (X :: XSYS) :- append XS YS XSYS.

The first rule says: appending YS to an empty list results in YS. The second rule says: appending YS to a list that has X as a head and XS as tails results in a list with X as a head and XSYS as a tail, when appending YS to XS results in XSYS.

It's a good exercise to try to convince yourself why the queries we saw above actually work, based on the small explanation I gave of how rules are executed. There really is not a lot of magic going on!

One might ask — why is it useful to have a language that relies on unification and relations instead of functions? One example is that when implementing a type checking procedure for a language, blurring the line between inputs and outputs allows us to get a type inferencing procedure essentially for free. But that is getting too much ahead of ourselves; we will see more in later posts.

With this out of the way, it is time to try our hand at writing our first predicate over the expressions we defined.

Writing an interpreter for our language

Let's go back to implementing our toy language now. Here is the base declaration of a predicate that relates an expression of our language with the value it will result in upon evaluation. We can use this predicate as an interpreter, if we give it a complete expression and a fully unknown value as arguments.

eval : (E: expr) (V: expr) -> prop.

Here's how we would use this, to evaluate/interpret a small example program:

eval (add (intconst 1) (intconst 2)) Value ?

Of course, this query fails at this point, as we have not given any kind of implementation for the eval predicate. So let's start with adding some rules to evaluate integer constants and integer addition:

eval (intconst I) (intconst I).
eval (add E1 E2) (intconst N) :-
  eval E1 (intconst N1),
  eval E2 (intconst N2),
  plus N1 N2 N.

(Note the distinction between add which is one of the constructors of expressions that we have defined, and plus, which is a built-in predicate for adding integers together.)

The first rule says: integer constants evaluate to themselves (because they are already values). The second one can be read as: the add expression evaluates to an integer constant N, when the two operands evaluate to the integer constants N1 and N2, and we also have N = N1 + N2. With these two rules, the query from above should now work:

eval (add (intconst 1) (intconst 2)) Value ?
>> Yes:
>> Value := intconst 3.

(Note that in a functional language when defining a function by pattern-matching, we have to give its full definition. In logic programming, we can add new rules for a predicate at any point of our program, similarly to how we can add new constructors at any point.)

Let's also add the cases for boolean constants, string constants and appending strings together. For the latter one, we can use the Makam builtin string predicate string.append:

eval (boolconst B) (boolconst B).
eval (stringconst S) (stringconst S).
eval (add E1 E2) (stringconst S) :-
  eval E1 (stringconst S1),
  eval E2 (stringconst S2),
  string.append S1 S2 S.

Let's try a couple more queries:

eval (add (stringconst "foo") (stringconst "bar")) V ?
eval (add (intconst 5) (stringconst "foo")) V ?
eval (add (stringconst "foo") (stringconst "bar")) V ?
>> Yes:
>> V := stringconst "foobar".

eval (add (intconst 5) (stringconst "foo")) V ?
>> Impossible.

Of course, the last query fails, as it should: we have only defined rules to handle the case where the operands to add evaluate to the same type of constant. That could be a deliberate choice depending on how we want evaluation in our language to behave.

How about arrays? For an array like [1 + 2, "foo" + "bar"], every member of the array needs to be evaluated. We can describe this using two rules:

eval (array []) (array []).

eval (array (HeadExpr :: TailExprs))
     (array (HeadVal :: TailVals)) :-
  eval HeadExpr HeadVal, eval (array TailExprs) (array TailVals).

Let's try this out:

eval (array [
       add (intconst 1) (intconst 2),
       add (stringconst "foo") (stringconst "bar")])
     Value ?
>> Yes:
>> Value := array [ intconst 3, stringconst "foobar" ].

(As an aside — we can do better than this. Remember when we said that Makam is a higher-order logic programming language? That means that we can define higher-order predicates — predicates that take other predicates as arguments — similarly to how we can define higher-order functions in a higher-order functional programming language. One example of such a predicate is map for lists, which is defined as follows in the Makam standard library:

map Pred [] [].
map Pred (X :: XS) (Y :: YS) :- Pred X Y, map Pred XS YS.

The evaluation rule for arrays would then be:

eval (array Exprs) (array Vals) :- map eval Exprs Vals.

More on this on a later installment.)

Evaluating records is a little more complicated. We need to evaluate the expressions contained within them, so that { foo: 1 + 1, bar: 2 + 2 } evaluates to { foo: 2, bar: 4 }. However, we also need to decide what to do about duplicate key entries, as in { foo: 1, foo: 2 }. For that, we will follow the JavaScript semantics for objects: duplicate entries for the same key are allowed, and the last occurrence of the same key is the one that gets picked — so the previous object evaluates to { foo: 2 }. We won't follow the JavaScript semantics when it comes to ordering the fields: instead, we will maintain the order of that appears in the source.

Let's see how to implement this in Makam. Here's a first attempt where we do not handle duplicate keys properly:

eval (record []) (record []).
eval (record (mkfield Key Expr :: Rest))
     (record (mkfield Key Value :: Rest')) :-
  eval Expr Value,
  eval (record Rest) (record Rest').

To account for duplicate keys, we need to split this last rule into two: one for the last occurrence of a key (where the key does not appear in subsequent fields) and one for any earlier occurrences. In this second case, the field can safely be ignored, as the language we are encoding does not have any side effects. To distinguish the two cases, we can use an auxiliary predicate contains_key (so a new prop), that succeeds whenever a key exists within a list of fields:

contains_key : (Fields: list field) (Key: string) -> prop.
contains_key (mkfield Key Expr :: Rest) Key.
contains_key (Field :: Rest) Key :-
  contains_key Rest Key.

And here's the rules for evaluation:

eval (record []) (record []).
eval (record (mkfield Key Expr :: Rest))
     (record Rest') :-
  contains_key Rest Key,
  eval (record Rest) (record Rest').
eval (record (mkfield Key Expr :: Rest))
     (record (mkfield Key Value :: Rest')) :-
  not(contains_key Rest Key),
  eval Expr Value, eval (record Rest) (record Rest').

Note the use of not here: basically, we are saying that this last rule applies whenever contains_key Rest Key is not successful1.

With these, the interpreter for our toy language is complete!

eval (record [
  mkfield "foo" (add (intconst 1) (intconst 1)),
  mkfield "bar" (array [ add (intconst 2) (intconst 2) ]),
  mkfield "foo" (add (intconst 4) (intconst 4))
]) V ?
>> Yes:
>> V := record [mkfield "bar" (array [intconst 4]), mkfield "foo" (intconst 8)].

Defining the concrete syntax for our language

One issue with our interpreter, which is quite evident in the query above, is that we have to use abstract syntax for writing down the terms of our language — and that's not always pleasant. Abstract syntax is often quite long-winded and verbose, even for simple terms. It would be nice to be able to use concrete syntax instead, to write queries like:

evalstring << { "foo": "a", "foo": [ "bar", 40 + 2 ] } >> Y ?

(The syntax form << .. >> is alternative syntax for strings in Makam, so that we don't have to escape the quotes " within it.)

What we need is a parser, that converts a string containing concrete syntax into the abstract syntax terms that we have defined. So we need to define a predicate with the type:

parse_expr : (Concrete: string) (Abstract: expr) -> prop.

Let's ruminate on this: given a query on parse_expr, what would happen if the second argument was a fully-known abstract syntax tree, whereas the first argument was fully unknown? In that case, we would be reconstructing the concrete syntax of an abstract syntax tree — namely, we would be using this predicate as a pretty-printer for our terms.2 So maybe parse_expr is not such a great name for our predicate, since we could use it both as a parser and a pretty-printer of expressions.

How about writing the predicate itself? Makam already has a syntax library that can help us implement syntax predicates like these by only giving a grammar for our language, similar to how parser generators are used in other languages. I will give an example of how to use it for the language we have defined in this post, and will just say that the parsing aspect of the library is based on PEG parsing3 and I am using an adaptation of Invertible Syntax Descriptions4 to the PEG setting so that the same grammar is used both for parsing and pretty printing.

Before looking at the code, let me briefly explain the components that go into it. First, we need to define "syntax constructors" which are akin to typed non-terminals in grammars: for example, a syntax constructor with the type syntax expr will be used as a handle that allows us to parse and pretty-print terms of type expr. Then, we need to give syntactic rules, which describe how to parse/pretty-print each term constructor (like stringconst, array, etc.). Last, we need to generate parsing/pretty-printing code for each toplevel syntax constructor; this is akin to running a parser generator to get the parsing code for our grammar. This step, just as all the other steps, happens within the same Makam program instead of requiring an external parser generator.

%open syntax.
baseexpr, expr : syntax expr.
field : syntax field.

`(syntax_rules <<

  expr ->
    add         { <baseexpr> "+" <expr> }
  / baseexpr ;

  baseexpr ->
    stringconst { <makam.string_literal> }
  / intconst    { <makam.int_literal> }
  / array       { "[" <list_sep (token ",") expr> "]" }
  / record      { "{" <list_sep (token ",") field> "}" }
  / { "(" <expr> ")" } ;

  field ->
    mkfield     { <makam.ident> ":" <expr> }
  / mkfield     { <makam.string_literal> ":" <expr> }

>>).
`( syntax.def_toplevel_js expr ).

Let's try out parsing and pretty-printing. We will use the syntax.run predicate, which does either of these two depending on its arguments:

syntax.run expr "{ foo: 1, bar: 2 + 2 }" Expr ?
syntax.run expr String (record [ mkfield "foo" (intconst 5) ]) ?
>> syntax.run expr "{ foo: 1, bar: 2 + 2 }" Expr ?
>> Yes:
>> Expr := record [ mkfield "foo" (intconst 1), mkfield "bar" (add (intconst 2) (intconst 2)) ].

>> syntax.run expr String (record [ mkfield "foo" (intconst 5) ]) ?
>> Yes:
>> String := "{ foo : 5 } ".

OK, let's unpack the code above a bit and explain what goes into it. The toplevel syntax constructor is expr, which we will use to parse and pretty-print expressions of our language. We also make use of two additional auxiliary syntaxes, one for base expressions and one for fields. expr represents the higher-precedence part of expressions — right now, this just stands for infix addition — while base expressions are the lower-precedence ones, which is everything else. The syntax library does not presently include any explicit support for describing precedence, and that's why we had to split into top-level and base expressions manually. Each syntax rule specifies the constructor that it is parsing/pretty-printing, along with any number of tokens and other syntax expressions that are needed. Each expression within the angle brackets needs to correspond to the type of each argument of the constructor: for example, in the first rule for the mkfield constructor, which requires a string followed by an expression, makam.ident is a syntax constructor of type syntax string and expr is of type syntax expr. The list_sep incantation for array and record is used to parse/pretty-print a list with the specified separator, which here is the "," token.

One thing to note about the workings of these rules is that contrary to context-free grammars, the rules here are applied in order and the choice is deterministic: given two rules like A / B, we attempt to parse/pretty-print using A, and only if that fails B is attempted. This has implications both for parsing and pretty-printing: for example, if we switch the order of rules for expr, we will never get to the rule for add when parsing, as baseexpr is already parseable on its own, and that's the first prerequisite for add. For pretty-printing, the order of rules means that a record will be printed as { foo: ... } instead of { "foo": ... } for keys that are identifiers; the string notation will be used otherwise. Another note is that left recursion is not permitted, hence we could not have a rule for add like <expr> "+" <expr> but need to break the recursion through baseexpr.

Last, one small note for the syntax of Makam itself and how these definitions actually work: the `( notation stands for a call to a staging predicate: that is, a predicate that generates further Makam code that is "inserted" in place. Here, syntax_rules transforms these grammar rules (given as a plain string with the notation << .. >>) into normal Makam rules that define parsing and pretty-printing; whereas syntax.def_toplevel_js generates some JavaScript code that is then inserted into a normal Makam predicate that will be used for parsing.

With syntax.run expr defined, we can now define a predicate that is more akin to the input-output portion of a REPL for our language, that takes an expression as a concrete string as input, evaluates it and returns the result as a string again:

evalstring : (ExprStr: string) (ValueStr: string) -> prop.
evalstring ExprStr ValueStr :-
  syntax.run expr ExprStr Expr,
  eval Expr Value,
  syntax.run expr ValueStr Value.

So what we do is that we first parse the concrete string; evaluate the expression to a value; and pretty-print the resulting value into concrete syntax.

We can now issue queries to try out our whole implementation so far. Note that this query block is editable, so you can try your own queries as well:

evalstring << { "foo": "a", "foo": [ "bar", 40 + 2 ] } >> X ?
evalstring << [ 1 + (12 + 12) ] >> X ?
>> evalstring << { "foo": "a", "foo": [ "bar", 40 + 2 ] } >> X ?
>> Yes:
>> X := "{ foo : [ \"bar\" , 42 ] } ".

>> evalstring << [ 1 + (12 + 12) ] >> X ?
>> Yes:
>> X := "[ 25 ] ".

Conclusion

We did cover quite a bit of stuff here: concrete and abstract syntax, the very basics of computation in logic programming, and writing an interpreter for a very simple language. Next time we will cover how to encode more complicated constructs, like functions, how to implement a type checker for our language in Makam, and more on the basics of higher-order logic programming and the Makam language design. So stay tuned!


  1. Negation in logic programming languages is a big topic, mostly because it breaks many of the invariants that hold about the language otherwise. For example: adding a new rule only makes more queries succeed, instead of making queries fail when they were previously succeeding; this does not hold in the presence of negation.

  2. The reality is more complicated of course — using the same predicate for both kinds of queries is not always possible for free, or will not always terminate. In a later post, we will explore this more in-depth.

  3. Bryan Ford. 2004. Parsing expression grammars: a recognition-based syntactic foundation. In Proceedings of the 31st ACM SIGPLAN-SIGACT symposium on Principles of programming languages (POPL '04). ACM, New York, NY, USA, 111-122. DOI: http://dx.doi.org/10.1145/964001.964011

  4. Tillmann Rendel and Klaus Ostermann. 2010. Invertible syntax descriptions: unifying parsing and pretty printing. In Proceedings of the third ACM Haskell symposium on Haskell (Haskell '10). ACM, New York, NY, USA, 1-12. DOI: https://doi.org/10.1145/1863523.1863525

Recent Posts