Skip to content
Branch: master
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
1598 lines (1190 sloc) 53.2 KB


I’m experimenting with taking an axiomatic approach to adding features to Lisp, using Arc as a starting point:

  • Containers for top level variables.

  • Source location tracking for profiling and showing the location of errors.

  • Automatically persisting a program’s data.

These experiments may not work out. The result may be too inefficient, impractical, or ugly to be useful.

While naturally a successful experiment is more exciting than a failed one, the failures may still be interesting: if something can’t be done within the constraints, why can’t it be done? What, exactly, is the trade-off in language design that turns out to be necessary?


The current release is 0.0.2.

  • Only parts of Arc are implemented so far.

  • As I’m focused on trying experiments, the code is very messy.


I push the implementation in one dimension until I can make no further progress without work on a different dimension.

For example, I have a sketch of an implementation (not yet included here) for automatically persisting a program’s data. Which is slow. Too slow to be usable. Optimization may make it faster, but it’s unlikely to ever be as fast as an existing language that doesn’t automatically persist data.

Slow though might be OK for high level scripting code, if there’s a way to run at full speed the parts of the program that don’t need to be persisted.

Maybe we add to the compiler a special form, perhaps we call it fast, which would compile the enclosed expression in “fast” mode which generated code which ran quickly but couldn’t be automatically persisted.

Now suppose we want to use a function defined in arc.arc like memo. memo maintains a cache of previous calls to a function and what it returned, so that if you call the function with the same arguments again you get the cached value. Would we like to be able to have the cache automatically persisted?

Persumably yes, we’d like to have that option, if we can. We could restart the program and still have the previously cached values. Which could be useful if the function is very expensive. (Suppose, for example, the function was making an API call to an external service).

On the other hand, if the function is only moderately expensive, and if we don’t particularly need to have the cache persisted, running a slow version of memo could easily swamp any performance improvements that we might get from caching the function call in the first place.

So we’d like to have a choice. To be able to use the standard, fast version of memo for when that makes sense, or the slower, enhanced version of memo when that would be useful.

Which kind of memo we get depends where it’s loaded.

(def memo (f)
  (with (cache (table) nilcache (table))

If memo is loaded in a place where table is the ordinary fast in-memory kind of table, then we get the standard, fast version of memo.

Or, if memo is loaded in a place where table is the slower, enhanced version of table which can automatically persist its data, we get the slower but enhanced version of memo.

I call these places “containers” because they contain the top level variables into which code can be loaded.

(Scheme uses the term “environment” and in Racket the analogous object is called a “namespace”).

Containers by themselves don’t implement a module system (though it might be possible to build a module system on top of containers), but containers also allow modification of how code loaded into them work, which isn’t something modules typically do.

Given that we want containers (or, at least I think we want containers), how should they work?

To take an axiomatic approach, I suggest that a container should be able to be treated as an ordinary Arc table. That you should be able to use an Arc table as a container if you wanted to. (obj + +) is a container that maps the top level variable + to the plus function.

Behind the scenes, when implementing Arc in Racket, we want to use a Racket namespace because that’s what Racket is optimized for.

In terms of the interface, how the container should appear in Arc, we want it to look like an Arc table.

That’s a hypothesis. An experiment. Containers might turn out to be a terrible idea, or a poor choice for allowing code to be compiled differently. So I explore along the container dimension to find out.

Meanwhile, if we do end up mixing slow code and fast code, it would be especially useful to have good profiler. So another dimension is source location tracking.

Hoped-for Features

(If the various experiments work out).

  • Source location tracking for profiling and to show the location of errors.

  • Different code can be loaded into different containers, where each container has its own set of top level variables (a.k.a. “global variables”).

    A container for top level variables can be an ordinary Arc table (e.g., for taking an axiomatic approach of implementing Arc in Arc), or, for greater efficiency when desired, a Racket namespace.

  • Top level variables, optional arguments, argument destructuring, and ssyntax are implemented as macros in Arc.

  • Includes fallintothis’ port of quasiquote from clisp, which correctly handles nested ,@ (useful for macro defining macros).

  • Multiple runtimes are supported simultaneously: for example, one container could implement Arc lists with Racket’s immutable pairs, while another container could implement Arc lists with Racket’s mutable mpairs.

  • Support for additional features such as block strings and writing and destructuring tables with {...}.


Amacx is licensed under the MIT license.

Libraries and components used by Amacx have their own license: Arc is licensed under the Artistic license, and the quasiquotation library ported from CLisp is licensed under the LGPL.

Running Amacx

First, you’ll need to bootstrap:

$ cd xcompile
$ racket -f expand.scm
$ cd ..

This only needs to be done once (unless you modify the source code of the compiler which is cross compiled), and creates an xboot subdirectory.

Try running Amacs:

$ ./amacx examples/hi.arc

hi.arc contains:

(use arc)

(prn "Hi")

(use x) means (load "x.arc") from your library path, unless "x.arc" has already been loaded before with use. Thus the (use arc) line loads “arc.arc” for you.

If you don’t include a program to run, you’ll get a REPL:

$ amacx

At this point though nothing but the builtins have been loaded yet ("arc.arc" isn’t loaded automatically).

$ amacx
> (prn "Hi")
Error: "prn: undefined;\n cannot reference an identifier before its definition\n  in module: top-level"

At the REPL you could first type (use arc) to load "arc.arc", or, the -u command line option will also do the same thing for you:

$ amacx -u arc
> (prn "Hi")

Amacx supports different runtimes (different implementations of the builtins), and you can use -r to specify which runtime to use:

$ amacx -r srcloc foo.arc

The default mpair runtime is similar to Arc3.2’s runtime (though it implements Arc lists using Racket’s mutable pairs), while the more experimental srcloc runtime allows forms read from a source file to be labeled with source location information.

-I path adds a directory to the search path that use loads features from. In a program you can also say (use lib "/path/to/library/dir").

-l file loads a file (unconditionally, whether or not its been loaded as a feature before, and without looking in the use include path). Thus if you’d like to load foo.arc and then get a REPL:

$ amacx -l foo.arc

You can also use --repl to get a REPL even if you’ve specified a program to run:

$ amacx --repl foo.arc

Command line arguments end up in argv:

$ amacx --repl foo.arc one two three
> argv
("one" "two" "three")

By default, amacx loads Racket’s errortrace module, which instruments the code to provide more information about run time errors at the cost of running the code more slowly. Use -n or --no-errortrace to not load errortrace.


  • Startup is slow.

  • errortrace is useful when used with the srcloc runtime which implements source location tracking to show the location of errors. It doesn’t help much with the mpair runtime (which, like Arc3.2, doesn’t track source locations). It’d be better to default to not loading errortrace when using runtimes that it doesn’t help with.

  • On an error in the mpair runtime the REPL avoids dumping the useless stack trace, but running a program with amacx program.arc currently does not.

  • At the moment we default to loading code into a symtab container (like an Arc table but with an error if an undefined top level variable is referenced); using a Racket namespace for the container is implemented (see src/racket-topvar.arc) but isn’t used yet.

  • The REPL supports command line editing (using Racket’s readline library) but not extended expressions that cross multiple input lines. (See e.g. pread.rkt for some of the messy details that are involved in doing this in Racket). Not sure yet how to incorporate this with the readtable hacking I’ve been doing.

  • You can invoke the amacx script from a different directory (/path/to/amacx/amacx foo.arc), but there isn’t a mechanism to copy the amacx script to another directory (such as your bin directory) and to run it from there.


You can install Amacx as a Racket “language” for use with DrRacket.

First, if you haven’t done it yet, do the bootstrap step described at the beginning of the previous section.

$ (cd xcompile; racket -f expand.scm)

Then change directory into rkt/amacx and run raco pkg install:

$ cd rkt/amacx
$ raco pkg install

This creates a “linked” package installation (not a copy) where Racket finds the package in this source directory.

Then in DrRacket you can load:

#lang amacx

(use arc)

(prn "hi")

Note that each #lang amacx file (module) is provided with its own separate container for Arc’s top level variables.

Currently the srcloc runtime is always used.


  • Installing the collection by copying isn’t supported: if we install from the root amacx source directory Racket tries to install everything and chokes… we’d either need to figure out a subset to give to Racket, or see if there’s some raco configuration to select what’s included in the collection.

  • We don’t have a way to specify the runtime here… perhaps use could be extended to be able to select the runtime? That could be tricky.

  • The REPL inside of DrRacket isn’t supported, we’d need a #%top-interaction… and I’m not sure if that could be made to work or not with the readtable hacking I’ve been doing to support block strings.

  • There’s no provision to export Arc functions back to Racket for using Arc to write Racket modules (like how Anarki can).

  • Startup is slow. And runs every time you click DrRacket’s "Run" button.

  • It’d be useful if Arc could ignore a #lang line… that way the same source code file could be loaded from both inside of DrRacket and by loading the program with amacx at the command line.

Introduction to Containers

A container is an Arc table which contains the values of your top level variables.

For example, if you run

(def foo () 123)
(= bar 456)

then foo and bar are added to the Arc table containing the top level variables with foo and bar as keys and a function and 456 as values.

Sometimes we call top level variables “global variables”, though they aren’t really “global” when we can more than one container.

A container can also be any object which can be made to act like an Arc table, such as a Racket namespace.

this-container provides a reference to the container that your code is being eval’d or loaded into.

> (def foo () 123)
> (foo)
> (this-container!foo)

We can create new containers:

> (= c (new-container))
> c!cons
> (c!cons 1 2)
(1 . 2)
> (use-feature c 'arc)
> (eval '(def foo () 123) c)
> c!foo
> (c!foo)

Here, new-container creates a new container and populates it with Arc’s builtins such as cons. The container however doesn’t have the def macro yet because Arc hasn’t been loaded. Next we load Arc, and then we can eval Arc code in the container.

Introduction to Runtimes

A runtime is how we implement a language once a program has been compiled and is now running.

For example, in Arc 3.2, cons creates an immutable Racket pair, car works with a Racket pair or a Racket symbol nil, and scar modifies an Arc cons with Racket’s unsafe-set-mcar!.

These are all runtime decisions. We could create a different runtime. For example, Arc’s cons could create a Racket mpair, Arc’s car could operate on mpairs, and scar could use Racket’s set-mcar!.

For Amacx I’ve written two runtimes so far. One I called “mpair” because it implements Arc lists using Racket’s mpairs. The other I call “srcloc” because it allows source location information to be attached to Arc forms such as lists and atoms to show where in the source an error occurred.

In a container, which runtime we’re using depends on which implementation of Arc’s builtins the container was populated with. For example, in a container set up for the “mpair” runtime, car will the version of car which operations on Racket mpairs. In “srcloc”, car is also able to be applied to Racket syntax objects so that it can be used with an Arc list read from a source code file.

Implementing Containers with Macros

When the Arc compiler encounters a variable, it knows whether it’s a lexical variable or not. (It keeps a list of lexical variables that are defined as it traverses down fn forms). Likewise, when it sees a variable that’s not a lexical variable, it knows that it’s a top level variable.

We could have the compiler translate a top level variable into a invocation of a topvar macro, giving us freedom on how we want to implement top level variables.

   (+ 2 3)
=> ((topvar +) 2 3)

Then, if for example we had an Arc table that we’re using as a container for top level variables:

arc> (= container (obj + + ...))
#hash((+ . #<procedure:+>) ...)

we might try implementing topvar something like this:

(mac topvar (var)
  `(container ',var))

Now the expansion would look like:

   (+ 2 3)
=> ((topvar +) 2 3)
=> ((container '+) 2 3)

We have a problem though! container as used here is itself a top level variable... so we’d have an infinite loop where container would get expanded into (topvar container).

When we generate code, we have the option of inserting either a symbol naming a variable which contains a value, or the value itself.

arc> (eval '(+ 2 3))
arc> (eval `(',+ 2 3))

What does ',+ produce? The , inserts the value of + (the + function), and the ' quotes it.

arc> `(',+ 2 3)
((quote #<procedure:+>) 2 3)

In most languages that have eval, only textual code can be evaluated and it’s not possible to evaluate code containing arbitrary values. Amacx is only possible because Racket has this feature.

Thus we can avoid the infinite loop problem this way:

(mac topvar (var)
  `(',container ',var))

and the expansion looks like:

   (+ 2 3)
=> ((topvar +) 2 3)
=> (('#hash((+ . #<procedure:+> ...)) '+) 2 3)

Purely for exposition, I'm going to use “⇒X” to mean “the value we got from X that we inserted into our code”. (“⇒” doesn’t actually appear in the language). Then I can show the expansion like this:

   (+ 2 3)
=> ((topvar +) 2 3)
=> (('⇒container '+) 2 3)

We’ve done something interesting here: we’ve implemented top level variables entirely with macros. The compiled code ends up having a reference to the container without the underlying runtime having to know about containers itself.

Arc3.2 compiles the Arc expression (+ 3 2) into the Racket expression (_+ 3 2). (The underline is used as a prefix to keep Arc names from colliding with Racket names). _+, in turn, is stored in a Racket namespace.

When compiling to Racket, it’s faster to reference top level variables this way than it is to look up variables in a hash table. We can optimize for Racket by checking whether our container is a Racket namespace:

(mac topvar (var)
  (if (isa container 'namespace)
       (ac-global-name var)
       `(',container ',var)))

where ac-global-name converts symbols like + into their prefixed version _+. Now we have an implementation that produces the same result as Arc3.2:

   (+ 2 3)
=> ((topvar +) 2 3)
=> (_+ 2 3)

(We’d also need to ensure that our compiler passed through prefixed variables rather than expanding them with topvar, or we’d get the infinite loop again).

Well... so at this point have we actually accomplished anything? Aside from doing what Arc3.2 does in a more complicated way?

I’m not entirely sure, but there are a couple ways that I think this might be useful:

First, when taking a axiomatic approach to language design, being able to use an ordinary Arc table as a container for top level variables strikes me, personally at least, as being rather intuitive. A specification, if you will, of how things should work. Or, at least, how they could work. (Whether actually useful or not, it was surprising to me that top level variables could be implemented using only macros).

And then in a particular situation (such as compiling to Racket), we can find an optimization (such as using a Racket namespace) which is more complicated but faster.

Second, we may not always be so lucky as to be implementing Arc on top of a language as powerful as Racket. While eventually we may implement something specialized for efficiency reasons like Racket does, when getting started it can be nice to have fewer things that need to be implemented to get off the ground.


  • In Arc3.2, setting an item to nil in a table removes that key from the table, and there isn’t a way to distinguish between an item being set to nil and the item not being present at all.

    Thus when using a standard Arc table for your container, referring to a top level variable that hasn’t been assigned yields nil instead of an error like “undefined: cannot reference an identifier before its definition”.

    I wondered if this might actually be a problem for me in practice. After all, how often do I misspell a function or macro name, or try to use one before its definition?

    It turns out, all the time :-) And then it’s a pain to have to track down where I’m trying to call nil.

    With source location tracking, this is less of a problem because we can see where the call to nil is being attempted. So this issue may be moot once source location tracking is fully implemented.

    In the meantime, it’s easy enough to have a specialized version of Arc’s tables which don’t remove items when they’re set to nil, and report an error when an undefined key is referenced. In the current code this is a called a “symtab” (a pun on “symbol table”, since in typical usage all the keys in a container are symbols).

  • Amacx currently boots using a symtab for the code’s container. Which kind of container to use should be an option.

Containers and Modules

Containers kind of look like modules. If other is a different container containing a function foo, and I want to “import” foo into my container, all I need to do is:

(= foo other!foo)

This doesn’t quite work for macros. If foo is a macro that expands into bar, I’d also need to import bar to get that to work.

There’s a trick though: rather than having foo expand into the symbol bar, I could have it expand into the value of bar (the function value if bar is a function, or the macro value if bar is a macro).

(mac foo ()
  (,bar "hi there"))

Now I can “import” foo and it will be referring to bar in its own container, and I don’t have to import bar as well for it to work.

Thus we can implement a module system using containers (without some other language extension such as namespacing symbols or using syntax objects), if we don’t mind using extra commas in our macro definitions. For example, here’s the Arc3.2 definition of complement:

(mac complement (f)
  (let g (uniq)
    `(fn ,g (no (apply ,f ,g)))))

If we want to be able to use complement without having to have compatible versions of no and apply (for whatever reason...), we can insert them too:

(mac complement (f)
  (let g (uniq)
    `(fn ,g (,no (,apply ,f ,g)))))

At first I found this awkward, but then I quickly got used to it. Now my brain sees a comma and says “ah, something we’re inserting” and it seems natural to me.

Just because we can do something doesn’t mean we have to. If you prefer, you could implement modules with macros by namespacing symbols (like Closure does), or by replacing symbols with syntax objects (like Racket does).

For myself, I find dealing with syntax objects or namespaced symbols overly complicated. I like being able to write macros with plain symbols, and, as it turns out, I don’t mind using extra commas to do so.

Amacx doesn’t implement a module system, though it makes it possible to implement a module and macro system using plain symbols (which most languages can’t do). But Amacx doesn’t prevent you from implementing a different kind of module system either.

As an analogy, consider how it’s easy to implement an object-oriented programming system in Arc if we want to, but Arc doesn’t need to implement object-oriented programming itself to let us do that.

Containerizing Code

There’s something that containers can do that modules don’t: we can use containers to do things to code, rather than simply using them to provide us with code like modules do.

As an example, here’s the definition of map1 in Arc3.2:

(def map1 (f xs)
  (if (no xs)
      (cons (f (car xs)) (map1 f (cdr xs)))))

What map1 produces, of course, depends on what cons produces.

Let’s say I wanted a version of map1 that produced, oh, I don’t know, hyperblocks. (I don’t know what a hyperblock is, but let’s pretend! :-)


(= other!cons hyperblock)
(load "map1.arc" other)

Thus containers are something we can load code into. We can load code into a container using load, or, if we have some kind of module system, we could import or otherwise copy code into a container as well.


Before we can start loading code into a container, we first need to populate it with the builtin functions, Arc’s primitives (or the primitives of your language that you’re using Amacx for).

Assuming we have a source to copy the builtins from, creating a new container with the builtins is as simple as making a copy:

(= my-container (copy my-builtins))

These builtins in turn define the runtime: for example, does cons return a Racket immutable pair or a mutable mpair? Does car only work on Arc lists, or does it also work on Racket lists (assuming they're different in the implementation)? Does ((eval '(fn args args)) 'a 'b 'c) produce an Arc list? The answers depend on what the builtins do.

At the moment there are two runtimes implemented in Amacx:

mpair strictly implements Arc lists with Racket mpair’s and nil, and Racket lists (constructed with Racket pairs and the Racket empty list) never appear in the Arc runtime.

For practical purposes less strictness is useful: it’s convenient if car also works on Racket lists for example. However I find mpair helpful in establishing a baseline for corner cases such as the behavior of using lists as a compound key in a hash table.

srcloc implements associating values such as symbols and lists with the source file locations that they were read from. Like mpair, cons produces a Racket mpair; but car and cdr can also be applied to Racket lists.

Other options are possible. We could have a runtime that implements Arc lists with Racket pairs munged with the “unsafe” car and cdr set operations like Arc3.2 does, or we could have a runtime where Arc lists are immutable by default if we wanted.

For my own Arc inspired language Eternal, I’d like to write a runtime implemented as an interpreter that can serialize closures. As a naive, unoptimized interpreter (at least at first), it would be very slow, but might still be useful for top-level, “scripting” code that can call faster code.

Since different containers can be populated with different builtins, we can have containers running different runtimes. Thus I could have one container running Eternal, and then for code that needed to be faster, that code could be loaded into an Arc container.


We can start amacx with a new container containing nothing but the builtins from a runtime:

$ amacx -r mpair
> 3

The container has builtins such as + (in this case, since I specified mpair for the runtime, mpair’s version of +), but since my container doesn’t even have topvar defined yet I’m not able refer to them directly:

> (+ 3 2)
Error: "need topvar macro defined for top level variable +"

I can load the topvar macro and then it will work:

> (use topvar)
> (+ 3 2)

Likewise, if Amacx implemented all of Arc, I could say (use arc strings) to load both Arc and Arc’s strings.arc library.

To be clear, this isn’t a module system. If it were, and Arc was a module, the Arc source code (arc.arc) would be loaded once, and then the definitions would be imported (copied) into whatever containers (or modules) were using Arc.

Instead, if we create a second container and use Arc in that container, the Arc source code gets loaded again. This is less efficient. However, it also means that the behavior of Arc can be modified by the container that it gets loaded into.

A second difference is that since code is loaded into the container, if A uses B and B uses C, A will see all the definitions in B and C. Nothing is hidden like in a module system where only public exports are visible.

On the other hand, if we want multiple containers and it’s too slow to load our code into each, containers are simply Arc tables. Arc’s copy will copy tables.

Within a container, use will only load features once. Thus any code that needs e.g. the strings library can say (use strings) without strings.arc getting loaded multiple times in the container.

The runtime can also provide features. Consider for example that apply has a basic form that applies a list to a function:

(apply + '(1 2 3))

For convenience, apply also allows you to supply additional arguments, which are also passed to the function:

(apply + 1 2 '(3 4))

In a host language like Racket it’s not hard to implement this more general form of apply which takes multiple arguments. (See for example ar-apply-args in Arc3.2’s ac.scm).

However in a less powerful host language it can be painful to implement even something as simple as ar-apply-args as a primitive…

What if apply happens to be called with thousands of arguments (perhaps as the result of code generation) but the host language doesn’t itself support proper tail calls? What if an exception is thrown from one of the arguments to apply, but the host language doesn’t itself support a compatible exception mechanism? What if continuations are involved and the host language doesn’t itself support continuations?

Implementing just the basic version of apply that takes the function and a single list argument can be challenging enough. Implementing the more general version of apply as a primitive might involve, in essence, compiling ar-apply-args by hand.

Thus we could prefer to have as a primitive that has to be implemented just the basic form of apply which takes a function and a single list argument (we could call that apply1), and then define the general form of apply in terms of the primitive:

(def apply-args (args)
  (if (no args)
      (no (cdr args))
       (car args)
       (cons (car args) (apply-args (cdr args)))))

(def apply (f . args)
  (apply1 f (apply-args args)))

On the other hand, Arc calls apply a lot, and if we’re using a more powerful host language like Racket where we can implement a more efficient version, we’d prefer to use that one.

What we can do is allow the runtime to also provide features. (Recall that use doesn’t load a source file for a feature if a feature has already been provided). One of our programs will use arc, which in turn will use apply. If the runtime implements apply as a primitive (like we can do in Racket), and provides apply as a feature, then (use apply) becomes a no-op. If the runtime doesn’t provide apply then the Arc version will be loaded.

The Arc version thus also serves as a specification. If there’s any confusion about how apply should work in the presence of, for example, Arc vs. Racket lists, exceptions, continuations, parameters, threads, or serialization… we can say it should work like the Arc version.

Source locations for errors

When reading source code from a file, we could use Racket’s read-syntax which will annotate the read forms with source locations: that, for example, this particular symbol foo was read from file foo.arc at line 6, column 10. Racket uses these source locations to show where an error occurred.

In Racket, these annotated forms are represented by syntax objects, which contain the annotated form, its source location if available, as well as potentially other information used by Racket’s macro system.

Consider a definition such as no in Arc:

(def no (x) (is x nil))

If this definition were read with Racket’s read-syntax, nil would not be the plain symbol nil, but instead would be a syntax object wrapping the symbol nil and containing a source location indicating that this nil was read at column 18. Thus calling this definition of no with a plain symbol nil (or a nil read from any other location) would return false.

In Racket, quoted forms are stripped of syntax information, which solves this particular problem. However in Arc macros can also manipulate forms read from a source code file, such here in Arc’s definition of withs:

(mac withs (parms . body)
  (if (no parms)
      `(do ,@body)
      `(let ,(car parms) ,(cadr parms)
         (withs ,(cddr parms) ,@body))))

If parms and body had been read from a source code file using Racket’s read-syntax, then no wouldn’t work (parms wouldn’t be plain symbol nil) and (car parms) and (cadr parms) wouldn’t work (parms would be a syntax object, not a pair that you can take the car or cdr of). Thus in Racket, macros need to be aware that they’re manipulating syntax objects instead of plain lists and symbols.

We can try a different approach. Imagine that when forms were read from a source file, they were labeled with source location information. Arc builtins that didn’t care about the source location information would ignore it, but it could be inspected when desired.

For example, (is 'foo 'foo) continues to return true because is is looking at the symbols foo without looking at the label indicating that the second foo came from a different column than the first one. Now macros such as withs don’t need to be modified because they’d continue to work.

Ideally we’d implement a general purpose labeling feature for Arc with an Arc representation of source file locations which could be used with any host language, and then for Racket translate the Arc representation into a Racket syntax object to support showing source location in stack traces.

For the current srcloc runtime I simply maintain the Racket syntax objects read using Racket’s read-syntax. Arc primitives such as is, car, and + unwrap syntax objects to obtain the wrapped form to operate on, ignoring the syntax information.

A few more steps are needed. The output of macros such as withs would lose source file location information because quasiquotation forms such as `(do ...) and `(let ...) produce new lists that aren’t labeled.

We can have the macro expander heuristically give the output of a macro expansion the same source file location information as the input, if the macro itself doesn’t itself return a form labeled with location information.

Here heuristic-loc gives an output expression e the same source location information as an input expression loc, unless e already has source location information (perhaps because a macro choose to explicitly return more precise location information):

(def heuristic-loc (loc e)
  (if (hasloc e) e (srcloc loc e)))

In this implementation we label forms with source location information by using Racket syntax objects; however syntax objects don’t survive being quoted. (In Racket (quote x) passes most objects xthrough unscathed, but syntax objects are converted to plain forms). Thus in thesrcloc` runtime we need to tunnel quoted values by wrapping them in a struct, and unwrapping them at run time when they’re used.

Finally, in Arc3.2, an Arc form such as (+ 3 2) is compiled into a Racket form such as (ar-funcall2 + 3 2), where ar-funcall2 checks to see whether the function is a Racket function that can be called directly or an Arc object that needs to be handled specially.

This makes the function application appear to originate from inside of the ar-funcall2 form. We can instead curry the function application and compile (+ 3 2) into the Racket form ((caller +) 2 3).

With these pieces in place DrRacket will show us a trace of where an error occurred:


  • The stack trace printed on error in the srcloc runtime is voluminous, messy, and most of the information printed is irrelevant. For example, the stack trace contains code excerpts… however, since what Racket sees is the Ail language output of Arc’s macro expansion, that’s what gets printed, not the original source code. And, many of the stack trace entries come from the runtime and don’t provide any useful information about what went wrong in the user code.

    Pruning the stack trace to show only the useful information would make it a lot more readable.

  • It’s not necessary to use Racket’s syntax objects to label forms with source location information in Arc; we could have some other mechanism to label objects in Arc. Once we got to the point of eval’ing the code in Racket we’d need to convert Arc’s representation of forms labeled with source location information into Racket’s syntax objects for Racket to be able to use them in error traces… but we also wouldn’t have to tunnel Arc’s representation through Racket’s quote.


In Racket we can bootstrap by loading Arc3.2, but if we want to implement Arc or our own language on top of a more primitive platform, it would be easier to cross-compile.

Thus we’d like to be able to write out the output of macro expansion (the macro expansion of our compiler implementation in particular) so that we can load it into a new runtime… including the symbols created by uniq.

Similarly, if we implement a runtime sometime that supports serialization, we’d again want an implementation of uniq which could create symbols that would survive being serialized. (We could also consider cross compiling as being a special case of serialization).

Traditionally Lisp implements unique symbols with “uninterned” symbols, that is, a symbol foo generated in one place won’t be equal to a symbol foo read somewhere else. However the identity of the symbol would be lost when written out to a file, so that wouldn’t work for cross-compiling.

Arc3.2 takes a quick-and-dirty approach of generating symbols by appending “gs” with an incrementing integer… good enough if we don’t have symbols named “gs” in our own code, but if we start loading macro expanded code created elsewhere how do we avoid collisions?

A solution is to generate a universally unique identifier with enough randomness that the likelihood of a collision is vanishingly small (even if we ran a computer the size of the universe for the lifetime of the universe).

> (uniq 'foo)
> (mac example ()
    (w/uniq foo
> (example)

An Arc implementation language

If we make fn a macro, then we can easily implement in Arc optional arguments and argument destructuring, which in Arc3.2 are implemented in the compiler.

In Racket, the primitive version of fn is called lambda, so our fn macro could perhaps look something like this:

(mac fn (args . body) (if (complex-args args) (complex-fn args body) `(lambda ,args ,@body)))

Here we check to see if the function arguments contain any optional arguments or argument destructuring; if they do we’ll generate code to handle that using complex-fn; if not we can simply delegate to Racket by turning the fn form into a lambda form.

A pleasant side effect is that argument destructuring now works with optional arguments. I don’t know if I’ll ever use use argument destructuring with optional arguments, but it’s nice that it works! :-)

A disadvantage is that now I’m reserving the name “lambda” for compiling fn into Racket. I can’t write my own function or macro that happens to be named “lambda” without messing up fn. I’d prefer that “lambda”, or whatever we’re using to compile functions, have some unique name that wouldn’t conflict with any of the names I might like to use in my own program.

In Racket, “lambda” doesn’t have to be called “lambda”; I could for example import lambda into a Racket module with a different name… for example if I wanted to call the primitive form of fn something like $fn. I’d need a Racket module that I can eval the Racket code the compiler produces, and in that module I can import lambda as $fn:

(require (only-in racket (lambda $fn)))

Still, whatever name I choose causes that name to be reserved. I don’t care what the name is, and I’d prefer it was a unique name that wouldn’t ever collide with a program I or someone else was writing. If I could use w/uniq with Racket code…

(w/uniq $fn
  `(module my-racket-module racket
     (require (only-in racket (lambda ,$fn)))

     ... and then in Arc ...

       (mac fn (args . body)
         (if (complex-args args)
              (complex-fn args body)
              `(,$fn ,args ,@body)))))

I can’t actually do that, of course, but if I could, $fn would end up being something like $fn--xVrP8JItk2Ot.

Someone could look at the macro expansion and see that it was named $fn--xVrP8JItk2Ot and would then be free to use $fn--xVrP8JItk2Ot themselves if they wanted to, but they wouldn’t collide with $fn--xVrP8JItk2Ot by accident.

Of course, we could pretend that we were using w/uniq, and output $fn--xVrP8JItk2Ot ourselves…

(mac fn (args . body)
  (if (complex-args args)
       (complex-fn args body)
       `($fn--xVrP8JItk2Ot ,args ,@body)))))

In essence, doing a bit of macro expansion by hand.

The primitive forms needed to implement Arc once macro expansion is complete are $quote, $fn, $assign, $if, and $call.

I call this primitive language “Ail” (for “Arc implementation language”). If a runtime can provide an implementation for Ail and a minimal set of builtins such as apply1, we’ll have a working implementation of Arc.

(Note that a “minimal set of builtins” includes implementing Scheme features such as proper tail calls, exception handling, continuations, and parameters… so that part may not be so easy).

The actual names of $quote and so on are $quote--xVrP8JItk2Ot etc., which allows someone the freedom to use $quote etc. as names in their own code without conflicting with the compiler.

This makes Ail a remarkably ugly language… yet the output of macro expansion usually is ugly, and no one cares because no one has to look at it.

Compiler varients

With the compiler written in Arc, we could extend the compiler simply by, for example, extending the definition of the compiler:

(defextend ac (s env) (is-ssyntax s)
  (ac (expand-ssyntax s) env))

This would offer a more powerful way of extending the language than plain macros. While beguiling, modifying the compiler in place turns out to be really painful when building a compiler to implement a new language.

With only a single compiler, the compiler ends up having to simultaneously correctly compile the language the compiler is written in (i.e. Arc) and the new language that the compiler is for.

Instead, my current approach is to first allow compiler rules to be specified and given names, and then a compiler can be constructed by supplying a list of compiler rule names to be used for that compiler.

For example, here’s a compiler for Arc:

(= a-compiler-for-arc
    '(nil-sym this-container lexvar topvar quote assign-lexvar
      assign-topvar fn if explicit-call macro implicit-call

The generated compiler can then be called with a container and an expression, and it’ll return the compiled expression.

Here’s an example of a compiler rule, one that recognizes a lexical variable and returns the variable itself as the compilation result:

(compiler-rule lexvar (and (isa e 'sym) (is-lexical context e))

The form of compiler-rule is (compiler-rule name test . body).

Both the test and the body is called within a function that has context and e as parameters. e is the expression being compiled and context contains the compilation context, including for example the list of lexical variables that have been defined at this point.

Inside of a compiler rule, expressions can be recursively compiled using compile:

(compiler-rule assign-lexvar (and (caris e '$assign)
                            (is-lexical context (cadr e)))
  (check-assign (cadr e))
  `($assign ,(cadr e) ,(compile context (caddr e))))

Here, $assign is the primitive version of Arc’s assign (which itself is an Arc macro that expands into $assign). With an expression like

($assign x (+ 2 3))

We pass through the variable name x unchanged, and recursively compile the expression (+ 2 3).

context contains a reference to the compiler generated by gen-compiler, so compile recursively compiles using the compiler that was originally invoked.


  • For the srcloc runtime, the source location heuristic is applied by compile. This seems overly specific (e.g. what if we wanted a compiler that didn’t apply source location heuristics, or wanted to determine source locations in a different way)… but probably better to wait for an example, as any speculative design I try to come up with now might not work for what would need to be done in practice.


With the compiler for Arc written in Arc, naturally we need a working version of Arc in order to run the compiler.

When using Racket, we can start with Arc3.2 (as its compiler is written in Racket) and so bootstrap by compiling our compiler for Arc using Arc3.2’s compiler.

For other target languages, an easy approach is to macro expand the compiler. The output of macro expansion is Ail, the primitive language implementing Arc after macro expansion. The target language then only needs to provide an implementation of Ail in order to run the compiler.

A gotcha here is that the output of macro expansion in Amacx isn’t purely textual, as it is in other languages. The output of macro expansion can include arbitrary objects such as procedures.

For example, implementing top level variables with containers using macro expansion means that the container is injected into the macro expansion output.

Similarly, if macros inject the functions they use, we end up with procedure objects in the output.

We could have a variant of our compiler and language definition in which macro expansion produces purely textual output (like Arc3.2 does for example). We could then use that version in the target language to bootstrap into the version which supports containers and the other features of Amacx.

There’s a trick we can do though. Although macro expansion allows injecting arbitrary values into the macro expansion output, the only values that actually do end up getting injected into the macro expansion output when we compile the compiler are the container and functions assigned to top level variables.

Thus as an additional step, we can convert the macro expansion of compiling the compiler into a purely textual form using the following algorithm:

  • If we see the container, replace it with a named value “this-container”.

  • If we see a procedure, do a reverse lookup in the container to find a key which maps to the procedure, and replace the procedure with that name.

For naming these values, I use a Racket boxed value for convenience. That detail doesn’t matter though, as we could use any textual representation that distinguishes named values.

As an example, consider the def macro which expands a function definition like (def caar (xs) ...) into Ail code like this:

   ($call ($quote ⇒sref) ($quote ⇒this-container) 'caar ($fn (xs) ...))

(Recall that “⇒X” is a notational convenience to mean “the value we got from X which we inserted into our code).

Here we call the sref builtin to set caar in the container to the function implementing caar.

Replacing the injected values with the named version results in a textual form that we can save to a file:

   ($call ($quote #&sref) ($quote #&this-container) 'caar ($fn (xs) ...))

(In Racket #& creates a box literal. Note however the textual representation isn’t all that important, we could for example generate JSON if we were targeting JavaScript… whatever would be convenient to read in the target language).

To bootstrap in a target environment, we apply the algorithm in reverse: replace a named “this-container” with the actual container we’re bootstrapping into, and replace a named function with the function value found in the container with that name.

Some of the named procedures will be builtins like cons, while others will be functions defined in Arc that we’re loading from the macro expansion output. Thus the algorithm needs to be applied one form at a time: first we replace named values with their actual values in the bootstrap container, then we execute the resulting Ail code; and then we proceed with the next form.

To get started with Amacx, first run the cross compiler which is written in Arc3.2:

$ cd xcompile
$ racket -f expand.scm

This creates the compiler.nail file, the Amacx compiler macro expanded into Ail with injected values converted to named values.

Then when we run Amacx:

$ amacx

it bootstraps by loading compiler.nail into a bootstrap container. This provides eval, load, use, repl, and so on, which Amacx can then use to load Arc code.

Supported features

While Amacx is a compiler, it should be able to support language features such as those described below, which in turn helps us see what capabilities the compiler should have.

By analogy, consider how a Linux device driver is independent of the Linux kernel (it can be developed separately and installed dynamically), but the kernel source tree includes device drivers as that helps make visible what the kernel needs to support for device drivers and allows device driver implementation to be refactored along with the kernel.

Curly brackets for tables

curly-table allows curly brackets to be used for table literals.

(use curly-table)

{a 1, b 2}

produces the same table as (obj a 1 b 2).

I found that having no syntax in the curly brackets and using e.g. {a 1 b 2 } without a comma was visually too confusing for me. Adding a comma to the syntax was enough, I found I didn’t need to, for example, also add a colon (like JavaScript) to make it visually discernible.

The comma needs to immediately follow the value expression to be recognized as a separator (instead of a quasiquote unquote expression).

The value expression can be omitted and defaults to being the same as the key (typically becoming variable reference), thus

(with (a 1 b 2)
  {a, b})

is like saying (with (a 1 b 2) (obj a a b b)) and results in the same table as (obj a 1 b 2).

With curly-destruct, tables can be destructed:

(use curly-destruct)

(let {a, b} (obj a 1 b 2)
  (list a b))

returns (1 2).

If you’d like to use a different variable name than what you’re using for the table key, the { var key } form can be used:

(let {a 'x, b 'y} (obj x 27 y 28)
  (list a b))

Interestingly, the object being destructed doesn’t have to be a table. The object is called with the key to obtain the value, and any callable object can be used (including a function):

(let {a} string

returns "a".


  • Implement a printer, and print tables with curly brackets.

Block strings

Racket, like many languages, implements a “here” string:

(prn #<<END

The characters after the #<< define the delimiter. The text on the following lines, up to the delimiter on a line by itself, becomes the contents of the string. Parsing continues after the line with the delimiter.

Using “here” strings can be awkward. Any indentation is included in the string, so if you don’t want the contents of the string to be indented you need to left-justify the string without indentation. Any part of the program which needs to appear after the string literal (such as the closing ) here in this example) needs to follow after the final delimiter.

Block strings are an alternative:

(prn §)
  │ hello
  │ there

The § (a section sign character) declares that the following lines will be a block string. Parsing continues after the §, so that here in this example the closing ) can appear naturally on the same line with the opening (prn.

A block string line is any number of spaces and tabs, followed by a (Unicode U+2502, the box drawings light vertical character), followed by exactly one space, and then the contents of the string literal. This allows the string to be indented naturally.

The resulting string literal is "hello\nthere", as if the code had been:

(prn "hello\nthere")

Newlines are used to separate lines in the string literal, but the last line doesn’t have a newline appended. (It’s a bit easier to add a final newline when you want one than to have to remove an unwanted newline).

The § is handled by a Racket readtable, so if the § appears inside of a string or as a character literal it doesn’t trigger parsing block strings:

(prn "§")
(prn #\§)


  • Block strings don’t fit in very well with Racket’s parsing model. In Racket, read is supposed to consume a single datum from an input port, potentially peeking ahead in the input port if it needs to, but not otherwise maintaining state.

    To allow Racket’s read to be used, I implemented a synthetic “splicing port” which wraps the input port. The splicing port presents to Racket the section sign, the characters on the line after the section sign, and the characters after the block string lines… but the block string lines themselves are excised from the input that Racket sees.

    This works, but it’s not very intuitive, and it means that in all the places where read might be used (at the REPL, loading from a file, and so on), the splicing port has to be set up explicitly.

    Though I haven’t worked on an implementation, I’m imagine that a more intuitive model might be to implement parsing using generators: the parser would generate datums and the program would consume them. This would also simplify ports (which are complicated in Racket) since with the parser able to maintain state, ports wouldn’t have to implement peeking.


  • Feature requests: if you might like to use Amacx as a compiler for your own Arc inspired language, please let me know what features it would need to be able to do that.

  • Unit tests: unit tests are vital for development. For a compiler like Amacx, they’re not just helpful and a time saver; unit tests makes development possible. (Some low hanging fruit here might be to port some unit tests from Anarki).

  • TODO’s: pick a TODO from here or in the source code and see if that’s something you might like to tackle! :-)

See for how to sign off on contributions.

Thank you!


You can directly accelerate the development of Amacx and Eternal with a donation:

Contact me

I can be reached at

You can’t perform that action at this time.